Triangular Matrix Multiply

Lowering to `dot_general` is not ideal in this case since it assumes dense tensors. We can pattern match triangular matrices and convert dot_general to a `enzymexla.trmm` which can be lowered to:

1. BLAS trmm with a custom_call (loop for the batched case)
2. cuBLAS trmm with a custom_call or to batched_trmm in the batched case
3. All other backends lower it back to a dot_general

We can possibly run the detection and lowering as a last pass, but any phase ordering issues needs to be confirmed once

Start by tracing the existing lowering to dot_general and identify where triangular matrix patterns could be detected. Confirm phase-ordering constraints before deciding where the conversion belongs. Done means triangular cases use enzymexla.trmm, with the listed BLAS, cuBLAS, batched, and fallback paths accounted for.

Written by the indexing model from the issue text.