Add more batched-MatMul broadcasting support and tests

[affifboudaoud]

This PR adds a few more tests and special case fixes for the Batched Matrix Multiplication implementation in DaCe. Related to PR #2180. The overall goal is to add full NumPy/PyTorch matmul broadcasting support, including multi-dimensional batch broadcasting, 1D vector operations, and accumulation modes.

Changes

Broadcasting Support

• 3D-4D broadcasting (5ffe768): Handle different batch dimension sizes (e.g., [b1, b2, m, k] @ [b2, k, n] = [b1, b2, m, n])
• 1D vector broadcasting (3a88e77): Support 1D inputs as row/column vectors (e.g., [k] @ [b, k, n], [b, m, k] @ [k])
• Added batch dimension alignment and indexing logic for partial broadcasting

Accumulation Support

• Beta parameter (bf97502): Fixed beta scaling for C = alpha*A@B + beta*C in batched matmul, GEMM, and GEMV
• Handles beta=0 (zero init), beta=1 (accumulate), and other values (scale then accumulate)

Test Files

• batched_matmul_test.py: Tests for all broadcasting cases
  • 3D standard batched matmul
  • 3D-3D broadcasting (different batch sizes)
  • 3D-4D, 4D-5D broadcasting
  • 1D-3D, 1D-4D vector broadcasting
• test_matmul_accumulate.py: Tests for beta parameter
  • Batched matmul accumulation
  • GEMM accumulation
  • GEMV accumulation