fix(transformer): enforce max_len in make_transformer_src_tgt_masks - #3086

Resolves #2344.

Problem
In make_transformer_src_tgt_masks, length_to_mask(abs_len) infers its sequence width from abs_len.max(). When a batch is padded to a fixed sequence length or aligned to a hardware boundary where wav_len.max() < 1.0, the resulting key-padding mask width is strictly smaller than src.shape[1], causing shape mismatch errors in downstream multi-head attention layers.

Solution
Explicitly pass max_len=src.shape[1] to length_to_mask when constructing src_key_padding_mask.

Added unit regression test verifying that mask dimensions remain aligned to the full padded sequence width when relative lengths are strictly less than 1.0. The test also executes MultiheadAttention with the generated mask.