[codex] DSv4: derive config from HF + legacy-load + exporter DTensor materialize

[Meirtz]

DRAFT — source-only split; needs a torch/GPU unit run before it is marked ready (I could only py_compile it offline).

Summary

The load/save-correctness half of the DSv4 load-train work, split out so it can land
without the parts that change forward numerics. Three files, all additive/correctness:

• config.py: derive compress_ratios from layer_types/compress_rates and
  num_hash_layers from mlp_layer_types; parse nested rope_parameters (main/compress);
  add hc_eps/rms_norm_eps fields and an untied-embeddings guard. (Current code leaves
  compress_ratios=[] / num_hash_layers=3 defaulted, which is wrong for real
  DeepSeek-V4-Flash configs.)
• checkpoint.py: accept legacy model-rooted HF key aliases (expert/attention/compressor)
  on load; fix the expert-name double-dot (mlp.experts..N → mlp.experts.N).
• ckpt/hf_weights.py: materialize DTensor params to local cpu-contiguous before
  safetensors save — avoids silent FSDP2-shard truncation on export.

Intentionally excluded (ship separately)

• mHC numeric change (hca.py/mhc.py + model call-sites) — changes DSv4 forward numerics
  vs the current golden; needs an HF-parity gate first (per the basic.constitution no-reference
  rule, same bar as fix(deepseek_v4): scale MTP aux-loss gradient via pre_forward_hook #56/ds4 (deepseek_v4): DSA indexer aux-loss scale hook + lift forced CP=1 (MTP/dense-CSA support CP>1) #59). Separate PR.
• dispatcher scatter_add_ hash-routing parity — correctness toward HF but changes hash-layer
  golden; separate PR + golden refresh.
• The combined static test file — to be carved to cover only config/checkpoint/exporter.

Validation status

• py_compile clean; verified no references to the excluded mHC/dispatcher code.
• Pending before un-draft: carve the config/checkpoint/exporter unit tests + one cluster
  unit run (toy DSv4 load/export round-trip).

[ISEEKYAN]

why do we need so many legacy** stuff