Merge upstream main (v2.3.0) into fork main

[Kaushikj-7]

Overview: Sync fork with state-spaces/mamba main (v2.3.0), keeping our fork-specific docs/configs and tooling.

Upstream highlights (not in fork): CI now builds arm wheels and publishes on failure; version bump to 2.3.0; deterministic backward kernel for mamba2 and new determinism utils/tests; selective_scan/triton kernels refreshed.

Fork-specific work retained: Dockerfile + .dockerignore; Colab training guide and replication guide; extra configs (125m, 1k/1m/65k); DNA training pipeline pieces (eval_metrics, dna_embeddings, mamba_wrapper, scheduler, train_mamba); hg38 download/sharding scripts; count_params helper; PDF paper copy.

[Kaushikj-7]

Merging upstream main (state-spaces/mamba v2.3.0) into our fork.

Highlights:

• Upstream: version bump to 2.3.0, deterministic backward kernel for mamba2, refreshed selective_scan/triton ops, CI builds arm wheels and publishes on failure.
• Fork: keep Dockerfile + .dockerignore, Colab training and replication guides, extra configs (125m, 1k/1m/65k), DNA training pipeline bits (eval_metrics, dna_embeddings, mamba_wrapper, scheduler, train_mamba), hg38 download/sharding scripts, count_params helper, paper PDF.

Conflict hot spots to resolve:

• mamba_ssm/ops/triton/ssd_* updates vs fork tweaks
• determinism tests vs fork removals
• src/training/train_mamba.py and related training helpers

After resolving conflicts:

• Run selective_scan/triton op tests and determinism suite
• Do a short training smoke test
• Rebuild Docker image to confirm still works