Minimax m3 weight free export

[amarquic]

Fixes weight-free (meta-device) torch.export/torch.onnx.export for MiniMax-M3's language decoder via QEFFAutoModelForImageTextToText. Each fix was
root-caused from a real export/compile failure surfaced by iteratively running examples/text_generation/minimax_m3_decode_only.py against the real
MiniMax-M3 checkpoint (and, for the final dtype-leak issue, a standalone repro harness that exercises the full 60-layer config directly —
weight-free export only needs the config at trace time, not real weights).

• checkpoint_key_resolver.py: resolve block_sparse_moe.e_score_correction_bias — the checkpoint stores this buffer one level shallower (directly on
  the MoE block) than the live module tree (nested under mlp.gate/block_sparse_moe.gate).
• modeling_minimax_m3_vl.py:
  • Avoid duplicate nn.Module registration of the shared MoEWeights instance (object.setattr instead of a plain attribute assignment), which
    otherwise made torch.export emit duplicate gate/up/down initializers under two FQNs — only one of which promote_initializers_and_build_spec would
    promote, leaving the other as an un-promoted meta tensor that fails at ONNX save().
  • Avoid device-bound lifted scalar constants in two shape[1] == torch.tensor(1, device=...) comparisons — the meta-device input made these
    dataless lifted constants that fail ONNX serialization.
  • Replace _qeff_minimax_clamp's torch.maximum/torch.minimum against a torch.tensor(..., device=hidden_states.device) constant with
    torch.clamp(min=, max=), which accepts plain Python scalars directly.
  • Cast hidden_states to lm_head.weight.dtype consistently at all four lm_head call sites (two were missing the cast the other two already had),
    fixing a float16/float32 Equal/matmul dtype mismatch the QAIC compiler rejected.
  • Cast the sigmoid router's top_k_weights back to the router's native (float16) dtype before returning. routing_weights is intentionally upcast to
    float32 for numerically-stable sigmoid scoring (matching upstream HF), but the derived top_k_weights was never cast back down, so moe_decode_bmm's
    down * topk_weights silently promoted the MoE block's output to float32. This was invisible with --num-layers 4 (the MoE layer was the last layer
    traced, and the lm_head cast mopped it up), but fatal on the real 60-layer model as soon as a later float16-weighted layer consumed the float32
    activation.
• blocked_attention_forwards.py: in blocked_kv_attention_forward_headpar_offline (the function MiniMax's blocking_mode="kv_headpar" dispatches to),
  compare start_index as a plain Python int rather than wrapping it in torch.tensor(start_index, device=query.device). The wrapped form traces to a
  device-bound lifted tensor placeholder that has no data under meta-device tracing; the plain-int comparison traces to a scalar-comparison op with
  the constant baked in as an attribute instead. Scoped to only this function, not the ~9 other occurrences of the same pattern elsewhere in the file
  that back other (non-MiniMax) blocking modes.