Use a plain parameter for high-precision grouped linear weights

[wujingyue]

Problem

GroupedLinear(single_grouped_weight=True) always registers the packed weight as a
GroupedTensor — a storage-less _make_wrapper_subclass whose payload lives in Python
attributes (rowwise_data, scale_inv, first_dims, ...).

That is necessary for quantized weights, whose grouped scale/amax/scale_inv buffers
must travel with the data. It is not necessary for high-precision weights.

GroupedTensorFromPyTorchGroupedTensor (csrc/type_converters.cpp) reads only four things
when there is no quantizer:

What the GEMM reads	Plain-tensor equivalent
num_tensors	w.size(0)
logical_shape	(G*M, N)
rowwise_data ptr + scalar_type()	w.data_ptr(), w.dtype
_with_gemm_swizzled_scales	constant False

Every other attribute is None and short-circuits. For uniform member shapes
make_grouped_tensor takes the branch where first_dims, last_dims and tensor_offsets
all stay None, so the descriptor carries no device-side metadata at all.

The wrapper is not free. It has none of the serialization plumbing QuantizedTensor carries
(untyped_storage(), the _to_copy handler, new_empty, __reduce_ex__), and it bans
reshape-family ops, so torch.save, torch.distributed.checkpoint and generic tensor ops do
not work on these parameters.

Change

Register high-precision grouped weights and biases as plain stacked (G, M, N) parameters,
and build a transient GroupedTensorStorage at GEMM call time.

This is already how the gradient side works — the wgrad path wraps a plain main_grad buffer
and hands autograd back a plain (G, M, N) tensor. This applies the same transient-wrap
pattern to the parameter. Quantized weights are unchanged.

Scope

Only the fusible op (ops/basic/grouped_linear.py) is converted. module/grouped_linear.py
keeps GroupedTensor: it passes per-GEMM members directly into _GroupedLinear.apply and
holds them via ctx.origin_weight_refs weakrefs, so member lifetime and autograd wiring
there need separate treatment.

Verification

H100 (CC 9.0), cuBLASLt 130501, NVTE_GROUPED_LINEAR_SINGLE_PARAM=1.

Confirmed the fused path actually engages and stays zero-copy:

bf16 graph-safe path supported: True
weight type: Parameter (3, 32, 64)
gemm descriptor: GroupedTensorStorage
zero-copy (same storage as param): True
device-side metadata: {'first_dims': None, 'last_dims': None, 'tensor_offsets': None}

New tests (6, all passing): representation, forward/backward parity against the per-group
layout, torch.save/load round-trip, and a guard that quantized weights still get a
GroupedTensor.

No regressions — running the unmodified test file against baseline vs. this branch gives
identical results:

Suite	Baseline	This branch
test_grouped_tensor.py (orig file)	20 failed, 161 passed, 35 skipped	20 failed, 161 passed, 35 skipped
test_fusible_ops.py -k grouped	50 failed, 137 passed, 223 skipped	50 failed, 137 passed, 223 skipped

Those pre-existing failures are an artifact of my test environment, not of this branch: the
container ships a .so compiled at e7c550c5 while the Python is main, so newer C++
symbols (e.g. tex.scaled_swiglu) are absent. They are identical on both sides.

black --line-length=100 --preview and pylint (10.00/10) clean.