This commit fixes #494.

This commit changes the type of the view size accepted by `nvgRenderFrame()` from `int` to `float` so it matches the type defined in `GLNVGcontext`. This would prevent a framebuffer missing a pixel in final rendering result.

- disable stroke AA gradient when not using AA

This commit updates font stash to render glyph's bitmap data on demand. This change is intended to fix the issue that `nvgTextBounds()` may return an unreliable result when font atlas is full. This situation usually happens when auto scaling a large font text to fit a small rectangle.

The `fons__getGlyph()` function now accepts a new parameter indicating whether bitmap data is optional, and only `nvgText()` requires it. All of `nvgTextGlyphPositions()`, `nvgTextBreakLines()`, `nvgTextBounds()` and `nvgTextBoxBounds()` functions are changed to work without glyph’s bitmap data.

Fixed return value of nvgText to be the position after the rendered
string, also fixed scaling of the value.

This commit adds a new `nvgShapeAntiAlias()` function for turning on/off antialias for specific shapes.

Previously, the Unicode range of the CJK chractaters only includes Chinese charactaters, Japanese Kanji, and CJK Symbols and Punctuation. The range are used by traditioanl Korean as well. This commit further includes Hiragana and Katakana in Japanese, and Hangul Syllables and Hangul Jamo in Korean to make CJK support more complete.

This commit fixes the bug that text may not render correctly when glyph is not found in fallback fonts.

This commit improves to display CJK characters mixed with English words so there is less spaces at line breaks.

This commit fixes weird line breaks when rendering CJK (Chinese, Japanese, Korean) characters. Each CJK character is more like a word instead of a letter. This commit fixes the logic of line breaks by introducing a new `NVG_CJK_CHAR` code point type and takes special care for formatting text including CJK characters.

This commit fixes the formats as suggested in #299 and fixes stack overflow when using `freetype` backend.

This commit updates the implmentation of global composite operation to follow the advices described at c4b865796d (r74248556)

This commit implements the `nvgGlobalCompositeOperation()` function to support blending between frames. All operations defined in HTML5 canvas API are supported. Also, it is possible to create custom composite operation by calling `nvgBlendFunc()` or `nvgBlendFuncSeparate()` functions.

This commit updates the `nvgBeginFrame()` function to forward the received `devicePixelRatio` parameter to the `renderViewport()` function so it is possible to configure the environemnt correctly when using `bgfx` backend.

- fixed stroke to use lino join style to calculate required vertex
count

Line break would duplicate a word if the first character of a word
would cause a line break

It guards addPoint from "stealing" the point from the previous path if they're the same, which makes it break when you do something like this:
```
moveTo(50, 50)
moveTo(50, 50)
lineTo(100, 100)
```

Before the fix there's no line shown. With the second moveTo, there's a second path added, but without the 50,50 point, because it was already added with the first moveTo. This makes the second path only contain the point 100,100 so it doesn't draw at all.

After the fix, the point equality check is ignored for the first point in the path, so the second path gets the 50,50 point too and that makes it behave as expected.

This is only one solution to the duplicated moveTo problem and it's likely not the best one, but it's one that's easy to add and it seems to work just fine.

This commit implements a new nvgCancelFrame() function for the purpose described in #168.

Currently, if calling nvgBeginFrame(), the nvgEndFrame() function must be called to finish the drawing. However, if knowing the drawing would take too much time and want to give up the drawing. There is no way to cancel it. nvgCancelFrame() gives another choice to end the frame without actually drawing.

- API changed!
- moved image repeat from nvgImagePattern() into image creation
nvgCreateImage*()
- made flip-y and premultiplied common image flags (not just GL)
- removed nvglImageFlags(), flags passed in via
nvglCreateImageFromHandle() flags
- nvgluCreateFramebuffer takes image flags as param

- removed duplicate call to nvgTextMetrics()

- set default lineHeight to 1.0

This prevents commands like `nvgQuadTo` and `nvgArcTo` from double-transforming the previous point when they pass it back to `nvg__appendCommands`.

- added function to allow to combine scissor rects using intersection

- added typedefs for all structs
- fixed some compiler warnings

- set stb_image flags so that it loads correct data in iPhone

Tweaked premake to remove some windows warnings.

- fixed stroke cap tessellation amount calculation

- added nvgQuadTo, draws quadratic bezier curve segment

- all colors in the shader is handled as premultiplied
- removed alphaBlend param from nvgBeginFrame(), it became obsolete
- added NVGL_TEXTURE_PREMULTIPLIED flag
- separated FBO example
- FBOs are treated as premultipied for correct alpha
- changed FBO storage from depth/stencil 24/8 to stencil 8
- moved discard to end of shader which restores the render speed again

- added (optional) 3-pass stencil trick to stroke rendering, reduces
overdraw artifacts, a little slower
- fixed some text bounds issues