[Quantization] Pin NVFP4 calibrator collect tests to reference path

After folding the Triton fast path into NVFP4MSECalibrator.collect(),
two tests in tests/gpu/torch/quantization/test_nvfp4_static_quantizer_cuda.py
broke because they inspect path-specific state:

- test_collect_and_compute_amax asserts ``cal._losses_sum is not None``
  with ``len == 126``. Only the reference 126-step sweep populates that
  list; the Triton fast path produces ``_best_amax_fast`` directly and
  leaves ``_losses_sum = None``.
- test_multiple_collections asserts that two ``collect()`` calls
  accumulate. The Triton fast path is one-shot by design and refuses
  a second collect until ``reset()``, so multi-collect is fundamentally
  reference-path semantics.

Fix: take the ``monkeypatch`` fixture in both tests and force
``MODELOPT_NVFP4_TRITON_SWEEP=0`` so they exercise the reference
accumulator. Triton-path coverage stays in
test_nvfp4_fp8_sweep_kernel.py (parity, dispatch predicate, end-to-end
mtq.quantize). The other tests in the same class (initialization,
candidate generation, per-block independence) are path-agnostic and
unchanged.

Signed-off-by: Chenjie Luo <chenjiel@nvidia.com>

Author: cjluo-nv

tests/gpu/torch/quantization/test_nvfp4_static_quantizer_cuda.py

@@ -164,8 +164,15 @@ def test_fp8_candidates_generation(self, device):
         assert torch.all(torch.isfinite(candidates))
         assert torch.all(candidates > 0)
 
-    def test_collect_and_compute_amax(self, device):
-        """Test collect and compute_amax workflow."""
+    def test_collect_and_compute_amax(self, device, monkeypatch):
+        """Test reference-path collect and compute_amax workflow.
+
+        Pinned to the reference 126-step sweep (``MODELOPT_NVFP4_TRITON_SWEEP=0``)
+        because this test inspects ``_losses_sum``, which only the reference path
+        populates; the Triton fast path produces ``_best_amax_fast`` directly and
+        is covered separately in ``test_nvfp4_fp8_sweep_kernel.py``.
+        """
+        monkeypatch.setenv("MODELOPT_NVFP4_TRITON_SWEEP", "0")
         num_blocks = 8
         block_size = 16
         per_block_amax = torch.ones(num_blocks, device=device)
@@ -193,8 +200,14 @@ def quant_func(x, amax):
         assert torch.all(torch.isfinite(amax))
         assert torch.all(amax > 0)
 
-    def test_multiple_collections(self, device):
-        """Test that multiple collections accumulate correctly."""
+    def test_multiple_collections(self, device, monkeypatch):
+        """Test that multiple collections accumulate correctly.
+
+        Multi-collect is reference-path-only — the Triton fast path is one-shot
+        and refuses a second ``collect()`` until ``reset()``. Forcing the env var
+        keeps this exercising the accumulator.
+        """
+        monkeypatch.setenv("MODELOPT_NVFP4_TRITON_SWEEP", "0")
         num_blocks = 4
         block_size = 16
         per_block_amax = torch.ones(num_blocks, device=device)