Improvements

[FL33TW00D]

• CLI for model conversions
• Extend wgpu-mm tiled matmul to support any input shape, then switch out our terrible kernels
• Improve matmul interface
• Distil whisper support
• Test error on quantized matmul and add superblock quant.
• Add GPU profiler
• Add test support to cargo-instruments for easy integration
• Fix god awful error handling
• Experiment: can you get from the wire to a wgpu:buffer with zero copies using a SharedArrayBuffer
• Extend allocator to use greedy by size improved
• Validate logit mutators
• Fix Operation / MetaOperation traits
• Inplace binary operations
• Prune dependencies and bifurcate them for platforms