feat(cuda-device): add f32 and f64 warp reduce utilities

[honeyspoon]

warp::redux_sync_* is integer-only. The surface is redux_sync_add/and/or/xor plus redux_sync_{min,max}_{i32,u32}, with no float form — which mirrors PTX, where redux.sync defines only integer operand types. It is a hardware limit, not a gap in the bindings.

The consequence is that anyone reducing floats across a warp writes the same butterfly shuffle by hand. This adds it once, for both precisions:

reduce_sum_f32   reduce_max_f32   reduce_min_f32
reduce_sum_f64   reduce_max_f64   reduce_min_f64

Each is five XOR shuffles at offsets 16, 8, 4, 2, 1, built on the existing shuffle_xor_f32 / shuffle_xor_f64.

Notes for review

Every lane ends with the reduced value, not just lane 0. That is deliberate — it lets a caller feed the result straight into a block-level reduction without a second barrier, and it is the property a block reduce built on top of these depends on.

Safe functions, not unsafe. These are warp-synchronous but take no pointers and require no convergence beyond what the underlying shuffle intrinsics already document, so there is no additional obligation to put behind unsafe.

#[must_use] on all six: the return value is the entire point, so discarding one is always a mistake.

Uses f32::max / f32::min for the max and min variants rather than an open-coded compare, so the NaN behaviour follows whatever those lower to rather than being independently specified here.

fmt --check clean, no warnings, no new dependencies.