Support tile+offset and tensor descriptors

[jansel]

Stacked PRs (oldest at bottom):

• Fix bug with unit sized dims and block_sizes #932
• Make HELION_FORCE_AUTOTUNE or kernel.autotune() skip the cache #930
• Add advanced compiler configurations #793
• Verify compiled kernels in subprocess #914
• Fix triton/torch.compile compability issue #927
• ->Support tile+offset and tensor descriptors #928

---

Support tile+offset and tensor descriptors

[jansel]

@v0i0 this PR should help with one of your issues by letting you write:

        for tile_n_ in hl.tile(0, N, block_size=block_n):
            tile_n = tile_n_ + tile_m.begin // M * N

rather than using the dynamic begin bounds.