[Pallas] Skip trivial reduction mask when RDIM size equals actual dim

[norx1991]

Summary

• When the Pallas backend sets RDIM_SIZE equal to the actual dimension size (no rounding), the reduction mask is always true and can be skipped entirely
• Previously, the mask was only skipped for power-of-2 dimensions — this worked for Triton (which rounds RDIM up to the next power of 2), but Pallas uses exact RDIM sizes, so the mask was never needed yet always generated for non-power-of-2 dims
• This change generalizes the check by comparing backend.static_rdim_size(numel) against numel, which naturally handles both backends
• Guards numel > 0 to preserve correct behavior for zero-size reductions (static_rdim_size(0) returns 1 via next_power_of_2, which would incorrectly skip the mask)

Before (unnecessary mask for Pallas on non-power-of-2 dim like 1000):

def _helion_pallas_reduce_non_pow2(x, out, _RDIM_SIZE_1: int):
    indices_1 = jnp.arange(0, _RDIM_SIZE_1, dtype=jnp.int32)
    mask_1 = indices_1 < 1000
    row = x[:, :]
    _mask_to = jnp.where(jnp.broadcast_to(mask_1[None, :], ...),
                         row, jnp.full([], float('-inf'), jnp.float32))
    max_val = ...jnp.max(_mask_to, axis=1)...
    v_1 = jnp.exp(v_0)
    _mask_to_1 = jnp.where(jnp.broadcast_to(mask_1[None, :], ...),
                           v_1, jnp.full([], 0, jnp.float32))
    sum_1 = ...jnp.sum(_mask_to_1, axis=1)...

After (no mask — clean and correct):

def _helion_pallas_reduce_non_pow2(x, out):
    row = x[:, :]
    max_val = ...jnp.max(row, axis=1)...
    v_1 = jnp.exp(v_0)
    sum_1 = ...jnp.sum(v_1, axis=1)...