Add Snake activation functor for EVT

[emre570]

Adds cutlass::epilogue::thread::Snake, a two-operand activation functor
implementing the Snake activation from Ziyin, Hartwig, & Ueda
(NeurIPS 2020, arXiv:2006.08195):

Snake_a(x) = x + (1/a) * sin²(a*x)

The per-channel learnable frequency a is passed as a second operand,
flowing through an EVT child (typically Sm90RowBroadcast). This enables
fused GEMM/Conv + Snake epilogues for neural vocoders, where Snake is the
standard activation in the WaveGen path. See issue #3141 for production
benchmarks (~2.1× median speedup on H100 vocoder shapes).

Implementation

• include/cutlass/epilogue/thread/activation.h — scalar Snake<T> and
  Snake<Array<T, N>> specializations, using cutlass::fast_sin for
  IEEE-precise device/host math.
• Slots into the generic variadic Sm90Compute primary; no new
  infrastructure needed.
• Follows house style: CUTLASS_HOST_DEVICE, kIsHeavy = true,
  CUTLASS_PRAGMA_UNROLL on the Array specialization.

Tests

Adds unit tests in test/unit/epilogue/thread/activation.cu:

• Epilogue_thread_snake.device_f32 — tolerance 1e-5
• Epilogue_thread_snake.device_bf16 — tolerance 2e-2

Both pass on SM90 (H100). Goldens are generated from a math.sin-based
float64 reference over 256 samples each of x, α, and expected output,
with α ∈ (0.1, 2.0) and x ~ N(0, 1).

Usage

// EVT tree: Snake(acc, alpha_broadcast)
using SnakeEpilogue = cutlass::epilogue::fusion::Sm90EVT<
    cutlass::epilogue::fusion::Sm90Compute<
        cutlass::epilogue::thread::Snake,
        ElementOut, ElementEpi,
        cutlass::FloatRoundStyle::round_to_nearest>,
    cutlass::epilogue::fusion::Sm90AccFetch,
    cutlass::epilogue::fusion::Sm90RowBroadcast<0, TileShape, float>
>;

Closes #3141

[emre570]

@hwu36 Hi sir, is it possible to take a look in your free time?