fix(cutedsl): preserve FP32 CE output precision

[justinhh4]

Summary

• force FP32 CE loss and z-loss stores through explicitly typed FP32 pointers
• prevent CuTe DSL 4.6 from narrowing feature-enabled FP32 training outputs to bf16
• preserve the existing single-kernel, in-place-gradient design

Problem

With nvidia-cutlass-dsl==4.6.0, FP32 HAS_GRAD specializations that enable z-loss, softcap, or label smoothing store the public loss outputs at bf16 precision even though the CE math and gradients remain FP32-accurate. For example, the z-loss test produced 8.7725 instead of the Triton/PyTorch result 8.8014; per-token values landed exactly on bf16 quantization points.

Rebuilding the loss and z-loss views with explicit Float32 pointers keeps those stores FP32. FP16/BF16 paths retain their existing typed stores. This does not add launches, buffers, or API changes, and FLCE is unchanged.

Validation

• python -m pytest --confcutdir=test/cutedsl test/cutedsl/test_cross_entropy.py -q — 158 passed, 1 environment-gated skip
• LIGER_KERNEL_IMPL=cutedsl python -m pytest --confcutdir=test/cutedsl test/cutedsl/test_cross_entropy.py::test_liger_kernel_impl_cutedsl_selects_cutedsl_ce -q — 1 passed
• ruff check src/liger_kernel/ops/cutedsl/ops/cross_entropy.py
• ruff format --check src/liger_kernel/ops/cutedsl/ops/cross_entropy.py