Hermetic Benchmarking Prep + CPU Optimizations

[Marco-Christiani]

• Polish some fair benchmarking task (mnist) for Zigrad vs Pytorch on CPU
• Support granular ns and allocation analysis for optimization
• Provide some comparison feature
• Make small optimization tweaks
• Everything should plan to run with a hermetic build chain for replication on clean hardware later and public verification (actually implement via hermetic build pr).

Tasks and Changes

• [x] Faster FlattenLayer (its still not what I wanted but just for timing)
• [x] Tracy for granular ns benchmarking and statistical analysis
• [x] Add support for gemm accumulation
• [x] Batch matmul with accumulation
  • [x] bmm accumulation forward
  • [x] bmm accumulation backward
  • [x] Tests
• [x] add more matmul tests for edge cases
• [x] Fair profiling scripts for torch and zigrad mnist
• [x] Plotting torch vs zigrad
• [x] Add lots log points, remove debug logging calls that had some expensive calculations
• [x] Faster LinearLayer (better bias handling)

Build modifications

• Optional tracy dependency

[Marco-Christiani]

The faster linear layer should be refactored slightly to support other shapes or just renamed to Flatten1D its hacky