AMD CI hit the same Memory access fault on the jiaguo/fp8-inductor-fusion-v2
branch (this PR) as on the original sgl-project#21734, because reshape() returns a
stride-preserving view when the new shape is compatible with the input
strides -- which (*q.shape[:-1], -1, head_dim) is for a QKV-split tensor.
The resulting non-contiguous tensor is then passed to ROCm's RMSNorm
kernel, which assumes contiguous inputs and faults.

Fix: keep the stride-preserving view only on the exact path that needs
inductor fusion (CUDA + piecewise_cuda_graph_compiler == "inductor"). On
ROCm and on CUDA with the eager PCG fallback, revert to the flat 2D
reshape(-1, head_dim), which is guaranteed to copy on a stride-tricked
input (the flat shape is not stride-viewable when the inner stride
spans H_total*D rather than H_q*D).

Also extract the reshape into a single `_reshape_for_qk_norm` helper,
replacing four inline copies in `apply_qk_norm`'s alt-stream and
straight-line branches.

Notes on semantics:
  - Original upstream `reshape(-1, head_dim)` happened to force a copy
    for the same stride reason, but only because the flat shape was
    un-viewable -- not by any explicit contiguity guarantee. We rely on
    the same property here for the ROCm path.
  - On the CUDA+inductor path, the stride-preserving view is what lets
    inductor fuse this reshape with the surrounding RMSNorm and FP8
    quant into a single triton kernel -- the original motivation of
    sgl-project#21734. Forcing .contiguous() would lose that fusion.