feat(ipc): optional GPU isolation for target-shard daemon (same-backend splits)

[davide221]

Summary

The target-shard IPC daemon is fork()+execv()'d and inherits the parent's environment with no GPU-visibility pinning (backend_ipc.cpp builds argv and execs; the daemon uses the raw --target-gpus indices). So the main process and the daemon both enumerate every GPU.

• Cross-backend split (CUDA main + HIP daemon) — harmless: the two runtimes see disjoint device sets.
• Same-backend split (e.g. HIP main + HIP daemon across two AMD GPUs) — two ROCr runtimes each initialize the other process's device at startup, which on some ROCm/gfx setups hard-faults the host.

This blocks same-backend multi-GPU target-shard splits for any model that uses the IPC path (qwen35 / gemma4 / laguna / deepseek4).

Change

Opt-in, backward-compatible:

• BackendIpcLaunchConfig gains child_env — extra env vars setenv'd in the child before exec (parent env untouched).
• When DFLASH_TARGET_SHARD_ISOLATE_GPUS is set, TargetShardIpcSession pins the daemon to only its assigned GPUs via {CUDA,ROCR}_VISIBLE_DEVICES and remaps --target-gpus to the 0-based pinned view.

Default (env unset) is byte-identical to today — child_env stays empty, the gpu list is unchanged. So existing CUDA+HIP splits are unaffected; there's no behavior change unless you opt in.

Why both CUDA_VISIBLE_DEVICES and ROCR_VISIBLE_DEVICES

Set both so the isolation is backend-agnostic: a HIP daemon honors ROCR_VISIBLE_DEVICES, a CUDA daemon honors CUDA_VISIBLE_DEVICES, and each ignores the other's var.

Validation

Validated in concept on a Strix Halo (gfx1151, iGPU) + external Radeon AI PRO R9700 (gfx1201, dGPU) HIP+HIP split. Applying the exact same pinning via an exec wrapper (ROCR_VISIBLE_DEVICES=<gpu>; exec backend_ipc_daemon) took the dual-GPU DeepSeek V4 Flash split from an instant host hard-fault to a stable, coherent generation across both GPUs (~8.4 tok/s, both GPUs computing). This PR moves that proven workaround into the launch mechanism.

Draft — testing checklist before merge

• Build (Linux HIP + Linux CUDA).
• In-code path on the HIP+HIP box (replace the exec wrapper with DFLASH_TARGET_SHARD_ISOLATE_GPUS=1).
• Regression: a CUDA+HIP split with the env unset (expect no change) and set (expect equivalent behavior + the daemon pinned to its GPU).
• Multi-GPU remote daemon (>1 remote GPU) index remap.