Warn on and normalize partial_rotary_factor declarations the DSpark modeling does not implement

[arthurgao2003]

Fixes #71.

Problem

The DSpark Qwen3 draft modeling applies RoPE across the full head dim (custom apply_rotary_pos_emb has no rotary/pass-through split, and Qwen3RotaryEmbedding yields full-dim cos/sin for this config), but build_draft_config starts from copy.deepcopy(target_config). A target — or a DFlash warm-start checkpoint — that declares partial_rotary_factor != 1.0 (Qwen3.5-family targets do) leaks the field into the saved draft checkpoint verbatim. The checkpoint is then self-inconsistent: weights trained with full-dim rotation, config declaring partial rotation.

Serving engines that honor the field, e.g. vLLM's dflash/dspark path, rotate head_dim * factor dims — a silent train/serve geometry mismatch. Measured on a Qwen3.5-27B-class target with a 5-layer draft warm-started from z-lab DFlash: offline AL 2.53 → vLLM serving AL ~1.4; recovers to 2.59 (matching offline) once the config declares 1.0. Full analysis in #71.

Change

• common.py: declared_partial_rotary_factor(config) (reads rope_parameters / rope_scaling / top-level) and check_partial_rotary_factor(config, normalize=False).
• qwen3/config.py: build_draft_config normalizes the declaration to 1.0 (with a warning) right after the deepcopy, so newly produced checkpoints describe the trained geometry.
• qwen3/modeling.py: Qwen3DSparkModel.__init__ warns when loading a config that still declares partial RoPE — covers pre-existing checkpoints and externally generated configs.

No numerics change: training already rotated the full head dim; this only makes the saved declaration truthful and surfaces the mismatch for old checkpoints.

Testing

• Helper unit tests: declaration read from rope_parameters / rope_scaling / top-level; normalize rewrites all present locations; clean configs stay silent and untouched.
• Integration: real Qwen3Config with partial_rotary_factor=0.25 through build_draft_config → one normalize warning, all declarations = 1.0; Qwen3DSparkModel init on the normalized config is silent, init on a dirty config fires exactly one warning; tiny model instantiates fine on CPU.