Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@tensorrt_llm/_torch/models/modeling_qwen3_next.py`:
- Around line 655-660: The code calls allgather(...) inside the
enable_lm_head_tp_in_adp branch (see enable_lm_head_tp_in_adp,
mapping_lm_head_tp, create_lm_head_tp_mapping) but allgather is not imported,
causing a NameError; fix by adding allgather to the import list from the
distributed module (the same place other distributed helpers are imported) so
that allgather is available when mapping_lm_head_tp and hidden_states are
all-gathered.

In `@tensorrt_llm/_torch/models/modeling_speculative.py`:
- Around line 798-800: The draft-mode code paths need the same Qwen3/Qwen3.5
handling as the single-engine path: update MTPDraftModel.__init__ to accept
"qwen3_5_moe_text" (in addition to "qwen3_next") and map it to the same MTP
class (Qwen3NextMTP imported from .modeling_qwen3_next), and extend
MTPDraftModelForCausalLM.load_weights to include the Qwen3/Qwen3.5 branch so the
matcher that handles qwen3_next also handles qwen3_5_moe_text; ensure any
spec_dec_mode.is_mtp_eagle() / two-engine checks treat both names symmetrically.

In `@tensorrt_llm/_torch/modules/mamba/gdn_mixer.py`:
- Around line 107-162: There is a duplicate Triton kernel definition named
fused_qkvzba_split_reshape_cat_kernel that overwrites the earlier one and causes
Ruff F811; remove the second definition (the entire function starting at the
later occurrence) so only the original fused_qkvzba_split_reshape_cat_kernel
remains; locate the duplicate by searching for the function name
fused_qkvzba_split_reshape_cat_kernel and delete the later block (including its
signature and body) to resolve the redefinition error and restore CI.

In `@tensorrt_llm/_torch/pyexecutor/_util.py`:
- Around line 1038-1049: The spec-layer extension currently appends spec entries
to hybrid_layer_mask/mamba_layer_mask without considering the caller-provided
layer_mask; update the logic so that if a layer_mask argument is provided you
first apply it to the existing hybrid_layer_mask and mamba_layer_mask (e.g.,
element-wise AND/zip) before computing num_layers and before extending with spec
layers, and ensure you align lengths (or pad/truncate) before extending with
get_num_spec_layers(spec_config) so the per-manager masks remain correct for
one-model separate draft KV mode.

In `@tensorrt_llm/_torch/pyexecutor/model_engine.py`:
- Around line 1525-1531: The slice assignment to inputs['position_ids'] using
previous_batch_tokens and previous_pos_id_offsets_cuda has closing bracket
indentation that triggers Flake8 E123; locate the block around
previous_batch_tokens > 0 in method/model where inputs['position_ids'] is
modified and reformat the expression so the bracketed index and the added value
align on the same indentation level (e.g., put the full slice [0,
num_ctx_tokens:num_ctx_tokens + previous_batch_tokens] on one line or keep the
opening bracket on the same column as the closing bracket) and ensure the
continuation of + (self.previous_pos_id_offsets_cuda[:previous_batch_tokens]) is
indented consistently; update the lines referencing previous_batch_tokens,
inputs['position_ids'], and previous_pos_id_offsets_cuda to satisfy E123.

In `@tensorrt_llm/_torch/speculative/mtp.py`:
- Around line 1038-1058: The reshape fails for padded 3D MRoPE batches because
the new branch in _select_mtp_position_ids/position_ids_gen assumes the last
axis is already truncated, but SpecDecOneEngineForCausalLM.forward still slices
position_ids as position_ids[:, :attn_metadata.num_tokens] (slicing the second
axis), leaving the time axis padded; fix by ensuring position_ids are truncated
on their last axis before this code runs — either update
SpecDecOneEngineForCausalLM.forward to slice position_ids on the final dimension
(e.g., position_ids[..., :attn_metadata.num_tokens]) or, inside the block that
computes position_ids_gen in _select_mtp_position_ids, explicitly slice/truncate
the last dimension to attn_metadata.num_tokens (or detect padded_num_tokens and
trim) before reshaping position_ids_gen and computing position_ids_delta so the
subsequent reshapes on position_ids_gen and position_ids_gen.flatten(...) cannot
encounter padded tokens.