Fix boolean tensor memory accounting

[Excelius-Wang]

dtype_byte_size(torch.bool) currently returns 1/8, but PyTorch boolean tensors use one byte per element. This undercounts boolean buffers in module-size estimates and can place tensors within a max_memory budget that is smaller than their actual storage.

Return 1 for torch.bool, reusing the existing sizing and placement paths. Keep the packed CustomDtype estimates unchanged.

A small randomly initialized GPT-2 using Transformers 4.57.6 has two 1 MiB boolean attention buffers. Before this change, Accelerate reports 511,496 bytes for a model with 2,346,504 bytes of parameters and buffers; after the change, the estimate matches. This demonstrates the storage-accounting issue, not a measured production OOM. GPT-2 in Transformers 5.15.1 no longer has these registered buffers.

Validation:

• New/updated regression selection: 15 failures and 8 passes before the fix; all pass after.
• Modeling utilities: 57 passed, 9 skipped (plus 4 passing subtests).
• Big modeling: 11 passed, 26 skipped on CPU.
• Covers empty and non-empty buffers, persistence, dtype overrides, nested totals, and automatic placement at budget boundaries.
• Full make quality and git diff --check pass.

Scope: this corrects boolean tensor accounting. The existing module-splitting path can separately omit a module's own buffers when no_split_module_classes is not supplied; this patch does not address that behavior or guarantee a hard runtime memory bound.