feat: support selector-local mixed quantization settings

[rdzhu225]

Summary

• add one-pass mixed quantization through repeated selector-local --select clauses
• reuse the existing activation-bits, strategy, group-size, scale-dtype, and chunk-size vocabulary per selector
• keep weight formats explicit (int4 / int8) so future fp4 / fp8 formats remain unambiguous
• export heterogeneous compressed-tensors groups with per-group pack-quantized / int-quantized formats and the official mixed-precision top-level format when required
• extend manifest generation, validation, DSV4 target classification, documentation, recipes, and end-to-end coverage

Interface

flagos-compressor quantize \
  --input /path/to/DeepSeek-V4-Flash \
  --output /path/to/DeepSeek-V4-Flash-Mixed \
  --select moe=int4 activation-bits=16 strategy=group group-size=32 \
  --select attention=int8 activation-bits=8 strategy=channel scale-dtype=fp32 \
  --n-candidates 8 \
  --backend cuda

The existing homogeneous interface remains unchanged:

--select moe --select attention \
--bits 4 --activation-bits 16 --strategy group --group-size 32

Selector-local and homogeneous selectors cannot be mixed in one command. In selector-local mode, activation-bits and strategy are required, so W8A16 groupwise, W8A16 per-channel, and W8A8 per-channel are distinguishable.

Scope

Selector-local mixed quantization currently uses the checkpoint-only MSE path. GPTQ, AWQ, and AutoRound are rejected explicitly in this mode. Runtime-specific DeepGEMM integration is intentionally outside this change.

Validation

• ruff check .
• local: 136 passed, 1 skipped
• H20 flagos_quant_infer / flagos: 137 passed
• real DSV4 W4A16 MoE + W8A8 attention dry-run: 33,924 groupwise INT4 actions, 241 channelwise W8A8 actions, 0 unmatched
• real DSV4 W8A16 granularity dry-run: 33,924 groupwise INT8 actions, 241 channelwise INT8 actions, 0 unmatched
• full DSV4 artifact validation: 46 shards, 103,109 tensors, 33,924 INT4 tensors, 241 INT8 tensors; all 44 wo_a weights stored as INT8