Fixes #2047

The native CPU dequantize_4bit mutated a 1-D shape to (1, n) before allocating the output, so even-length 1-D inputs dequantized to 2-D. The (1, n) view is only needed for the C++ kernel's 2-D indexing, and m = prod(shape[:-1]) is already 1 for 1-D inputs — this PR drops the mutation so the output keeps the caller-provided shape, matching the CUDA backend, the default/MPS backends, the registered fake kernel, and the CPU backend's own odd-length fallback path.

• Output values are bit-for-bit identical before and after over 36 configurations (fp32/fp16/bf16 × nf4/fp4 × blocksize 64/128 × n 128/256/4096); only the shape changes, (1, n) -> (n,).
• New regression test test_dequantize_4bit_1d_shape (24 params): fails 12/12 on CPU without the fix, passes 24/24 (cpu+mps) with it.
• Test4bitBlockwiseQuantOps: 536 passed, 24 skipped, 60 xfailed. test_functional.py -k 4bit: 1958 passed. Full pre-commit suite passes on touched files.
• Verified on macOS arm64 (NEON path) + MPS; the fix is backend-dispatch Python only, no csrc change. I could not run the CUDA/AVX-512 paths on this machine, but they are untouched.

Disclosure: prepared with AI assistance (Claude Code); I reviewed the change and take responsibility for it.