[Pallas] Fix TPU min_dot_size for matmul autotuning

[norx1991]

Summary

• Re-land [pallas-tpu] fix default configs for TPU examples #1731 (reverted in Revert "[pallas-tpu] fix default configs for TPU examples (#1731)" #1740 due to test_loops.py crash, which no longer reproduces on the current codebase)
• Set TPU _min_dot_size to (8, 128, 128) matching the Mosaic MXU tile dimensions (M>=8, K>=128, N>=128)
• Restructure device-type checks in _min_dot_size to be explicit (tpu, xpu, cuda) instead of fallback-based
• Add test_matmul_default to exercise autotuned matmul without hardcoded block_sizes
• Clamp min_dot_size to tensor dimensions on Pallas backend to prevent blocks exceeding tensor size
• Add test_matmul_smaller_than_min_dot_size to verify matmul works when dimensions are smaller than min_dot_size
• Add direct regression test verifying config_spec min_size constraints are [8, 128, 128] on TPU
• Remove xfail for test_reshape_input_types on Pallas (unblocked by the clamping fix)

Why correct min_dot_size matters for matmul autotuning performance

_min_dot_size feeds into default_config(), which generates the initial configuration that the autotuner uses as its starting point. It also constrains the autotuner search space via update_min_block() in matmul_ops.py. With the wrong default (16, 16, 16), the autotuner explores configs with K=16 or N=16 tiles that get zero-padded up to 128 before pl.dot, wasting autotuning time on suboptimal candidates.

With (8, 128, 128), the search space is pruned upfront so every trial uses MXU-compatible tiles, improving both autotuning speed and the quality of the resulting configs.

Clamping for small tensors

When a tensor dimension is smaller than min_dot_size (e.g., a 16x16 matmul with min N=128), update_min_block() would force the block size to 128 — larger than the tensor. Triton handles this via automatic out-of-bounds masking, but Pallas BlockSpecs can't. The fix clamps min_size to size_hint (the tensor dimension) on the Pallas backend, letting the dot-level padding in matmul_utils.py handle the MXU requirement at codegen time instead.