Add TIR Schedule and TIR support for consumer fusion

[guillon]

Motivation

TVM's legacy TE scheduling API is no longer available in recent TVM releases. Add an opt-in TIR scheduling path while preserving the existing TE path and XTC scheduling interface.

Description

This introduces schedulable/scheduled TVM expression abstractions that lower TE expressions to either TE schedules or TIR PrimFunc schedules. The TVM scheduler emits TIR transformations for tiling, interchange, caching, buffering, unrolling, vectorization, parallelization, and producer/consumer fusion. It also adapts unroll/vector combinations to TIR code-generation constraints and supports padded pack_at caches through storage alignment. Compilation, IR dumping, C emission, and library generation handle both schedule forms.

The TIR path is selected with the backend tir_schedule option or the exploration CLI's --tir-schedule flag. LoopNest records and displays producer and consumer fusion annotations. FileCheck coverage includes scheduled matmul and fused producer/consumer graphs, while retaining TE fusion coverage.

Commits

• Fix schedule description dump expectations.
• Add TE-to-TIR lowering and TIR scheduling support.
• Add fuse_producer_at to LoopNest.
• Emit producer fusion for TVM TIR schedules.
• Add the fuse_consumer_at scheduling primitive and TVM TIR support.
• Adapt unrolling and vectorization for valid TIR code generation.
• Add padded pack_at support for TIR schedules.

Discussion

The TIR path is opt-in; TE scheduling remains the default.

[codecov]

Codecov Report

❌ Patch coverage is 92.90780% with 20 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
src/xtc/backends/tvm/TVMScheduler.py	93.43%	3 Missing and 6 partials ⚠️
src/xtc/backends/tvm/TVMOpsCompiler.py	93.40%	3 Missing and 3 partials ⚠️
src/xtc/backends/tvm/TVMCompiler.py	84.84%	4 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!