Bridge+Memory leak fix+More functions

[vipuldivyanshu92]

Added Bridge APIs
Added temp workaround for 119 ANE compile limit
Offloaded the Classifier & Softmax to the ANE

[maderix]

Review & Benchmark Report

Thoroughly reviewed and benchmarked this PR. Great contribution — thanks for putting this together!

Benchmark Results (Stories110M, 30 steps)

Metric	Baseline (train_large)	This PR (train_large_ane)	Improvement
ms/step (train)	108.6	91.1	-16% (17.5ms saved)
Compiles per restart	72	99	+27 more kernels
Compile time	3.7s	4.1s	+400ms
Loss trajectory	10.39→9.78	10.39→9.78	Identical ✅

Where the gains come from

• Classifier forward: 1.3ms ANE vs ~9.3ms CPU cblas — ~8ms saved
• Softmax: 2.7ms ANE vs ~75ms CPU — ~7ms saved
• RMSNorm backward: 0.12ms/eval on ANE — marginal gain but good to have

Tests

• test_rmsnorm_bwd: PASS ✅ (max error 0.009, well within 0.05 threshold)
• test_classifier: All 4 subtests PASS ✅ (rmsnorm, classifier fwd, softmax, classifier bwd)

Correctness

Loss trajectories match baseline exactly — the ANE-offloaded ops produce numerically equivalent results.

Notes for follow-up

• Compile budget is tight (99/~100 limit) — may want to optimize
• gen_classifier_bwd() is unused (CPU cblas is faster for that shape) — could remove or document as experimental
• The bridge library is a nice abstraction for Python/ctypes consumers

Thanks for the contribution! 🎉