[codex] Fix DSv4 DCP checkpoint placements for DTensor-like params

[Meirtz]

Summary

This PR fixes the MLite DCP fallback path for DTensor-like parameters whose
checkpoint placement differs from the wrapped parameter placement.

For DSv4 expert weights under EP, the checkpoint protocol supplies the expert
mesh and placements, but the previous helper rebuilt the DCP tensor from the
FSDP2 parameter mesh/placements. In the reproduced failure this let rank 0 load
rank 1's expert shard for layers.3.mlp.experts.fc1.weight0.

Changes

• Use checkpoint mesh/placements when wrapping DTensor-like local shards for
  real sharded checkpoint placements.
• Preserve the parameter mesh/placements for unsharded checkpoint placements,
  so empty local DTensor shards do not force full materialization.
• Compute checkpoint global shape and contiguous stride from checkpoint
  placements when needed.
• Treat hc_head.* as replicated before the generic DSv4 head placement
  rule.
• Add targeted unit coverage with fake DTensor-like params for checkpoint
  placements, multi-axis sharding, unsharded checkpoint behavior, and empty
  local save behavior.

Validation

Local checks:

git diff --check
PYTHONDONTWRITEBYTECODE=1 /Users/lmei/anaconda3/envs/cua/bin/python -m py_compile \
  experimental/lite/megatron/lite/primitive/ckpt/dcp.py \
  experimental/lite/megatron/lite/model/deepseek_v4/lite/checkpoint.py \
  experimental/lite/tests/unit/primitive/test_training_checkpoint.py

• Direct no-pytest helper harness passed in the same cua Python/Torch
  environment for checkpoint-placement DTensor wrapping, multi-axis checkpoint
  shard shape expansion, and unsharded checkpoint behavior that preserves the
  parameter mesh.
• 2026-06-22 refresh: targeted pytest collection in the cua environment was
  attempted, but the file skipped before collection because local
  megatron.core.dist_checkpointing imports require triton, which is not
  installed. The no-pytest helper harness was rerun and passed for checkpoint
  placements, multi-axis shape expansion, unsharded empty-local save behavior,
  and empty local copy no-op.

GPU evidence from existing run artifacts:

• Pre-fix peer-shard mismatch reproduced on GB300, H100x2, and B100x2.
• Post-fix B100 ep2 SAVE_LOAD=1: analysis-save-load-ep2-b100-fix5.json
  reports overall=smoke_pass with all rank max deltas at 0.0.
• Post-fix B100 ep4 SAVE_LOAD=1: analysis-save-load-ep4-b100-fix5.json
  reports overall=smoke_pass with all rank max deltas at 0.0.
• Post-fix H100x2 ep2 SAVE_LOAD=1: job 1346071,
  analysis-save-load-ep2-h100-fix6.json reports overall=smoke_pass, and
  rank 0/rank 1 save_load.comparison.max_delta=0.0.
• Supplemental post-fix H100x2 pp2 SAVE_LOAD=1: job 1346142 passed with
  finite train metrics and rank 0/rank 1 save_load.comparison.max_delta=0.0.
  Local evidence is recorded from the launcher terminal stream at
  runs/20260620-mlite-next-gates-pr-packaging/remote-results-dsv4_dcp_pp2_save_load-terminal/results/metrics.json.
• Scoped DCP continuity, B100x2 ep2, MTP_ENABLE=0: job 1346239 passed
  save-load-continue versus uninterrupted training. loaded_step=1; rank 0
  compared 112 local tensors and rank 1 compared 111 local tensors; both ranks
  reported comparison.max_delta=0.0, no missing keys, and no shape
  mismatches. Evidence:
  runs/20260621-mlite-dsv4-dcp-continuity/remote-results-attempt2/results/metrics.json.
• Scoped DCP continuity, H100x2 ep2, MTP_ENABLE=0: job 1346254 passed the
  same save-load-continue gate. loaded_step=1; rank 0 compared 112 local
  tensors and rank 1 compared 111 local tensors; both ranks reported
  comparison.max_delta=0.0, no missing keys, and no shape mismatches.
  Evidence:
  runs/20260621-mlite-dsv4-dcp-continuity/remote-results-h100-attempt3/results/metrics.json.

Boundaries

This PR does not claim:

• full DSv4 Flash support,
• default FlashMLA/fused attention support,
• DSv4 MTP train continuity; the continuity evidence above disables MTP to
  isolate DCP from a known MTP/mHC DTensor/Tensor blocker outside this PR,
• GLM5.2, VERL, Qwen, or LoRA support.

The full local pytest suite was not run locally. Targeted pytest collection now
starts in the cua environment, but the local environment is still missing the
triton dependency needed by megatron.core.dist_checkpointing.