Fix misleading GPU probe diagnostics and add a GPU offload toggle (GPU.7, GPU.8)

[NCCU-Schultz-Lab]

Two fixes found by actually running the M-GPU validation recipe on a local CUDA
box (RTX 5060 Ti, CUDA 13.0) — the code has been GPU-code-complete but unvalidated
since session 54.

GPU.7 — a broken CUDA install was reported as "gpu4pyscf not installed"

gpu4pyscf-cuda13x does not declare the NVIDIA CUDA math libraries as
dependencies, so a by-the-book install left import gpu4pyscf failing with
ImportError: Failure finding "libnvJitLink.so". QuantUI reported that as:

GPU offload not available
  reason: gpu4pyscf not installed (see README → 'Optional: GPU acceleration')

The package was installed. The message sent the user back to an install step
they had already completed, with the real cause discarded.

Cause: ModuleNotFoundError is a subclass of ImportError, so
except ImportError cannot tell "the package is absent" from "the package is
present but its import chain is broken". Both landed in the same branch — and that
branch logged nothing, so quantui log tail couldn't recover the cause either,
despite the docstring promising it would.

Now:

• ModuleNotFoundError → "not installed", with the right pip install hint.
• Any other ImportError → "installed but failed to import: {exc}", including the
  missing library name, plus a pointer that CUDA libraries are the usual cause.
• Both paths log; the broken-chain case logs at WARNING.
• probe_gpu() -> (available, name, reason) is now the single source of truth.
  The CLI previously re-implemented the probe to build its own reason string,
  which is precisely how it drifted from what the dispatcher actually decided. The
  Status badge reads the same triple.

is_gpu_available() keeps its exact 2-tuple contract; the cache moved onto the
shared probe and cache_clear / cache_info are forwarded, so existing callers
and tests are unaffected.

GPU.8 — warn when a detected GPU probably won't help

Quantum-chemistry SCF is FP64 throughout. Consumer and workstation GPUs gate
double precision to roughly 1/32–1/64 of their single-precision rate, so offload
on such a card can be slower than the CPU. Measured on this box (4096² DGEMM):

	float32	float64
CPU (20 cores)	1 550 GFLOP/s	792
RTX 5060 Ti	15 346 GFLOP/s	346

Real B3LYP single points bore that out — 0.44×–0.91× of CPU wall time across
benzene/def2-SVP, benzene/def2-TZVP and ibuprofen/def2-TZVP. Before this change,
any detected CUDA device was presented as unqualified good news.

• is_low_fp64_device(name) classifies by device name against the known
  strong-FP64 datacenter families. Unknown names are classified low-FP64
  deliberately: consumer hardware is the common student case, and a spurious
  advisory costs far less than a silent halving of throughput.
• Advisory shown in the Status tab and by quantui gpu check.
• Offload still auto-engages. We warn rather than override: FP64 weakness is a
  tendency, not a guarantee, and a name-based auto-skip would silently overrule
  the user on hardware the heuristic got wrong.
• New persistent toggle — Status tab → Settings → "Use GPU when available",
  stored as compute.gpu_enabled. Additive settings section, schema version
  unchanged, unreadable settings never disable the GPU.
  QUANTUI_DISABLE_GPU=1 still wins over it, so scripted benchmarks are
  unaffected.

The toggle clears the detection cache on change — without that it would appear to
do nothing until restart, since the probe is cached for the process lifetime.

Testing

• New tests/test_gpu_diagnostics.py — 28 tests: the ImportError/ModuleNotFoundError
  split (including the exact libnvJitLink.so failure), logging, the preserved
  is_gpu_available contract and its cache-clear behavior, the settings gate, and
  FP64 classification across real device names. Platform-independent; no GPU, no CUDA.
• Two existing CLI tests were rewritten, deliberately.
  test_reports_missing_gpu4pyscf asserted that a bare ImportError produced
  "gpu4pyscf not installed" — it had encoded the bug. It now uses
  ModuleNotFoundError, and a new test locks the broken-chain case so the
  regression can't return. test_happy_path_when_gpu_detected was patching
  is_gpu_available; it now patches probe_gpu.
• Verified on the real GPU: the advisory fires on the RTX 5060 Ti; toggling the
  setting off makes an actual B3LYP run return gpu_used=False; toggling it back
  on restores gpu_used=True.
• Full suite: 1882 passed, 17 skipped. pre-commit clean at the pinned
  versions.

Not in this PR

The M-GPU speedup exit criterion still needs an H200 — it cannot be met on a
consumer card for the FP64 reason above. Everything upstream of it (detection,
migration, numerical agreement to 8 dp, gpu_used → perf_log → dashboard
pairing) is now validated locally, so that H200 session should be short.

🤖 Generated with Claude Code