Improve debug-mode diagnostics for out-of-bounds tile indexing

[daedalus5]

Description

When tile code indexes outside a tile's shape, Warp should report a clear,
coordinate-specific message before failing in debug mode. The diagnostic should
identify the tile storage kind, the offending coordinate dimension, the invalid
index value, and the valid range, for example:

Warp tile index out of bounds in shared tile: coordinate dimension 0 has index 1, outside valid range [0, 1) (tile rank=1)

This should cover shared tile indexing and register tile indexing paths that can
otherwise fail with generic assertions or CUDA error messages.

Context

The motivation is to make tile OOB bugs easier for users to debug. In a CUDA
failure case, an invalid shared-tile access can surface later as a generic CUDA
failure such as device-side assert triggered, without enough information to
identify which tile coordinate was invalid.

In debug mode, Warp has enough shape and coordinate information at the tile
layout layer to produce a much more useful message before the assertion. This
does not need to change release-mode behavior; the goal is better debugging
output when users enable debug mode.

[daedalus5]

Completed with 5df5550

[shi-eric]

Progress update: 5df5550 landed debug-mode diagnostics for out-of-bounds tile indexing.

The change adds debug-only bounds checks in the tile layout layer before the existing assertion path. When an invalid tile coordinate is used, Warp now reports the tile storage kind, offending coordinate dimension, invalid index value, valid range, and tile rank. This covers shared tile indexing and the register tile path requested in the issue; shared linear indexing also reports the invalid linear index and valid range.

Test coverage was added in warp/tests/tile/test_tile_oob.py:

• test_shared_tile_oob_reports_tile_index checks shared tile positive OOB diagnostics on CPU.
• test_shared_tile_negative_oob_reports_tile_index checks shared tile negative-index diagnostics on CPU.
• test_register_tile_oob_reports_tile_index checks register tile diagnostics on CPU.
• test_shared_tile_2d_oob_reports_dimension checks that multi-dimensional shared tiles report the offending dimension.
• test_cuda_shared_tile_oob_reports_tile_index checks the CUDA shared-tile path emits the Warp diagnostic before the device-side assertion.

TestTileOOB was also added to the default and debug unittest suites. The issue can stay closed as completed.