[gpu_cache] Only rasterize & upload after queue has successfully fit in cache

See merge request !125

Allows faster failing without wasted rasterization & upload. Produces a 
1.16-1.29 speed up in cache_bad_cases benchmarks
Update changelog

Add CacheBuilder::rebuild, Cache::to_builder

Closes #122

See merge request !124

Add resizing benchmark

Move gpu_cache benchmarks to benches/ & add cache thrashing benchmark

See merge request !123

Wrap gpu_cache tests in a module

Don't use default-features on dependencies

See merge request !122

This reduces the compile time, ensures that depending crates don't get
features forced onto them that they don't need / want / can't have and
works around a cargo bug around features.

Update approx -> 0.3

See merge request !121

Update dev-dependencies & examples

See merge request !120

Add some notes clarifying the top left origin of glyphs

See merge request !119

Swap fnv for fx hashing in gpu-cache hashmaps (#119)

See merge request !118

Provided by rustc-hash crate ~30% gpu-cache speed boost
Minor refactor to rasterizer code

Improve lazy_static -> SharedBytes conversion

Remove internal hashing in rasterizer

Use fnv hashing in rasterizer

Small key hash speed improves draw benchmarks by 9-65%

Add reference tests for drawing

Mirror the reference tests in function with fast byte-insert callbacks
verified by blake2 hash.

Reference tests draw a glyph and compare with a reference "correct"
render. Should protect against regression when modifying draw logic.

Update readme + remove doc version ref