…ta-pytorch#1802)

Summary:

## Problem

CUDA RNG produces different random sequences on different GPU architectures (e.g. V100 vs A100 vs H100) even with the same seed set via `torch.manual_seed()`. This causes stochastic gate CUDA tests to be flaky in CI — the same test passes on one GPU type but fails on another because expected values were hardcoded for a specific architecture's RNG output.

Additionally, `test_p_norm_decay` uses exact `assert ==` for floating-point tensor comparison, which fails on GPU due to floating-point precision differences.

## Solution

**CPU-seeded RNG approach**: In CUDA test subclasses, patch `_sample_gate_values` to generate random noise on CPU (where `torch.manual_seed` is deterministic across all hardware) and then move the tensor to the GPU device. This keeps the full training codepath exercised (noise + mu → clamp → gather → multiply) while ensuring cross-architecture determinism.

For `LazyGaussianStochasticGates`, both `initialize_parameters` (mu initialization) and `_sample_gate_values` (noise sampling) happen on-device after `.to(cuda)`, so both are patched to use CPU RNG.

Since CUDA tests now produce identical values to CPU tests, the `if cpu / elif cuda` branches in base test files are removed, along with associated `pyre-fixme[61]` comments.

For `test_p_norm_decay`, exact `assert ==` is replaced with `assertTensorAlmostEqual` with `delta=0.01` tolerance.

## Files Changed

- `test_gaussian_stochastic_gates_cuda.py`: Patch `_sample_gate_values` with CPU-seeded `normal_()` sampling
- `test_kuma_stochastic_gates_cuda.py`: Patch `_sample_gate_values` with CPU-seeded `uniform_()` sampling + Kumaraswamy transform
- `test_lazy_gaussian_stochastic_gates_cuda.py`: Patch both `initialize_parameters` and `_sample_gate_values`
- `test_gaussian_stochastic_gates.py`: Remove cpu/cuda branches (4 tests)
- `test_kuma_stochastic_gates.py`: Remove cpu/cuda branches (4 tests)
- `test_lazy_gaussian_stochastic_gates.py`: Remove cpu/cuda branches (12 tests)
- `test_p_norm_decay.py`: Use `assertTensorAlmostEqual` instead of exact equality (2 tests)

Differential Revision: D97775614