fix(sparse_mla): avoid immutable load-coordinate rebindings

[ErvinXie]

Summary

• give the V3.2 scheduled decode consumer0, consumer1, and producer paths distinct load-coordinate names
• remove TileLang eager builder warnings about immutable ldg_tx and ldg_ty values being re-bound
• preserve coordinate expressions, memory indices, and control flow

Context

This is the organization-owned mirror of MooreThreads#2 so the fix can be reviewed and integrated independently while the upstream PR is pending.

TileLang lowers these local expressions as immutable SSA-style bindings. The V3.2 kernel reused the same Python names in three mutually exclusive thread-role branches, but the eager builder sees the enclosing IfFrame as active and emits four warnings per kernel construction.

Validation

• AST alpha-equivalence against upstream main after normalizing renamed identifiers
• python3 -m py_compile mate/sparse_mla/tilelang/sparse_mla_v32_decode_fwd_scheduled.py
• pre-commit run -a

MUSA runtime pytest was not run on the development host because MUSA torch and tilelang are unavailable there.