fix(rccl): Extend fp8 device arithmetic optimizations and restore unroll 8 in gfx1250 device codegen

[KawtharShafie]

Restore unroll 8 in gfx1250 device codegen so FP8 kernels link and launch (RCCL_UNROLL_FACTOR=3). Extend gfx950 FP8 device optimizations to gfx1250

Motivation

FP8 ReduceScatter on gfx1250 fails on rccl/gfx1250 after #8880 narrowed device codegen to unroll 16/32 only. FP8 Ring/Simple requires the unroll-8 kernel variant and fails with invalid device function. gfx1250 was also excluded from FP8 device optimizations in rccl_float8.h.

Technical Details

• generate.py: Build unroll 8, 16, and 32 for gfx1250 (unroll 8 required for FP8 launch and RCCL_UNROLL_FACTOR=3; 32 remains the default per commSetUnrollFactor).
• rccl_float8.h: Extend gfx950 FP8 cast/upcast optimizations to gfx1250; use gfx942-style hadd intrinsics (gfx950 scalef32 is unsupported on gfx1250).

Related: #8880 — 4ee4eb0

Issue Tracking

JIRA ID: ROCM-286794

Test Plan

• Build RCCL for gfx1250 (./install.sh -l)
• Run reduce_scatter_perf @ 64M, MPI 4 ranks, Ring/Simple
• Compare FP32 and FP8 with default unroll and RCCL_UNROLL_FACTOR=3 (unroll 8)
• Compare against rccl/gfx1250 baseline (c318fa3) and prior good build (e0ebb75)

Test Result

Build	FP8 Ring/Simple	FP8 busbw (unroll 8)
c318fa3 (baseline)	FAIL (invalid device function)	...
This PR	PASS	...
e0ebb75 (reference)	PASS	...

FP32 Ring/Simple behavior is unchanged from the baseline. FP8 Ring/Simple no longer fails on the default path and completes successfully. With RCCL_UNROLL_FACTOR=3 (unroll 8) and applying the fp8 device optimizations to gfx1250, FP8 bandwidth improves over both the broken baseline and the prior reference build.

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/rocm-systems/blob/develop/CONTRIBUTING.md.