[Research] Preserve Qwen3.8 Flash Next full-compile checkpoint

[Leonccaa]

Purpose

Preserve the validated Qwen3.8 Flash Next full-compile path as a fork-internal
research checkpoint. This PR is intentionally based on
feat/qwen38-flash-next-sm70-core, so its diff contains only the compile-graph
work and does not expand the upstream model-support PR.

This is a correctness/fallback checkpoint, not a throughput claim and not an
upstream merge candidate as a single patch.

Validated behavior

• Four Tesla V100 32 GiB GPUs, TP4, FP16 activations, W4A16 NVFP4/Marlin, PLE
  CPU offload.
• One dynamic compile range (1, 256) compiled successfully.
• Mixed prefill/decode captured PIECEWISE and batch-one decode captured FULL.
• CUDA Graph capture added about 0.05 GiB per GPU.
• The 8-token correctness gate matched the current eager oracle exactly.
• The 64-interval benchmark matched the prior no-compile output hash in all
  three repeats:
  261ac6d4ac7500646010f2ae56882b3d0d020ce71f5279de2db1747c54f7cee3.

Performance result

Path	Mean decode	Median decode
no-compile FULL_DECODE_ONLY	26.433 tok/s	26.427 tok/s
full compile + FULL decode graph	25.803 tok/s	25.735 tok/s

The full-compile route was 2.38% slower by mean throughput. The whole-decoder
opaque boundary restores correctness but prevents Inductor from optimizing
inside the layer, so this patch should remain a reference/fallback while the
opaque scope is reduced incrementally.

Patch structure

• Dynamic-range warmup no longer treats a size-one capture as a safe warmup
  for a multi-size range.
• PLE CPU-offload wait, runtime slice, and optional FP8 dequantization are
  kept behind one compile-safe custom op.
• Qwen4Exp decoder, input embedding, and final HyperConnection mixer use
  caller-owned-output custom-op boundaries for exact compiled execution.
• Functionalization restores the original output/cache tensors for those
  caller-owned operations.

Checks

• Compile-range warmup tests: 5 passed.
• New PLE compile/offload tests: 3 passed.
• Targeted pre-commit, including Ruff and mypy: passed.

Upstream boundary

Potential upstream changes should be split and justified independently:

1. The generic dynamic-range size-one warmup fix.
2. The Qwen4Exp PLE runtime-shape/offload synchronization fix.

The whole-layer opaque wrappers and their functionalization rules stay in this
fork-only research PR until they are narrowed and demonstrate a performance or
maintainability benefit.