cuda: backport upstream revert of "reset cuda context after reading memory size" (#24715)

[min23asdw]

Close #237

### Suggested fix

Only reset a device that has never been used. Add an ever_active flag to
ggml_backend_cuda_device_context , set it at the three sites that do active_count++
(buffer alloc, host buffer alloc, backend init), and gate the reset on it:

if (ctx->active_count == 0 && !ctx->ever_active) {
CUDA_CHECK(cudaDeviceReset());
}

This preserves the original intent -- avoiding a lazily created context that permanently
holds VRAM when memory is queried before anything has been allocated -- while never
resetting a context whose kernel state other code still assumes is live.

- [x] Before the fix, llama-bench --n-cpu-moe 32,34,36,38,40,42,44 (6 model reloads)
aborts with CUDA error: invalid argument right after the first configuration
- [x] With the fix, the same sweep completes all 14 rows with no abort

Why a revert rather than a targeted fix

The first commit in this PR (f3c29b7) gated the reset behind an ever_active flag so a
device that had already been used would never be reset. It works, but upstream has
already settled this differently:

• upstream reverted 0f7fada56 entirely in ea21e0395 (Revert "cuda: reset cuda context after reading memory size (#23935)" ggml-org/llama.cpp#24715), authored by the same
  person who wrote the original commit
• upstream considered and closed without merging cuda: bump context generation on device reset to re-apply kernel attributes and fix a crash during model reload ggml-org/llama.cpp#24682, which kept the reset and
  re-applied kernel attributes via a context generation counter -- the same class of fix
  as ever_active
• the reset caused three separate problems upstream, all now closed:
  • CUDA: ggml_cuda_compute_forward: MUL_MAT failed ggml-org/llama.cpp#24328 CUDA: ggml_cuda_compute_forward: MUL_MAT failed (the crash reported here)
  • cuda: reset cuda context (#23935) causes multi-GPU split buffer crash during model load ggml-org/llama.cpp#24374 multi-GPU split buffer crash during model load
  • [Bug/Performance] cuda: reset cuda context (#23935) causes huge delays during model load with "--fit on" (common_fit_params: takes 20x longer) ggml-org/llama.cpp#24608 --fit model load taking 20x longer

This branch has diverged from upstream at this point: it carries 0f7fada56 but not the
revert. The second commit (af8ac64) cherry-picks ea21e0395, which supersedes the
first commit and brings this file back in line with upstream.

• Before: llama-bench --n-cpu-moe 32,34,36,38,40,42,44 -ngl 99 -ctk turbo4 -ctv turbo3 -b 4096 -ub 2048 aborts with CUDA error: invalid argument
  immediately after the first configuration
• After: the same sweep (7 configurations, 6 model reloads) completes all 14 rows
  with no abort
• Release build with -DGGML_CUDA=ON, no errors

[min23asdw]

I double check in ggml-org it already know this issue
https://github.com/ggml-org/llama.cpp/pull/24682 -> CLOSED
the fix ggml-org#24328 <--- same like i do.
ref
ggml-org#24374 (crash during model load)
ggml-org#24608 (slow 20x)

the ggml-org pull/24682 they not fix they revert
ggml-org#23935
ggml-org#24715

[min23asdw]

test [OK] - model reload (7 configs, 6 reloads) for single GPU CUDA sm_89

not test - multi-GPU

[TheTom]

Confirmed the second commit is byte-identical to upstream's revert ea21e03 (ggml-org#24715), and agree the revert beats the ever_active gate given upstream closed ggml-org#24682 in favor of reverting. Post-merge test on GB10: llama-bench --n-cpu-moe 0,8,16 -ctk turbo4 -ctv turbo3 (2 model reloads) completes with no 'CUDA error: invalid argument'. Nice writeup of the upstream history — made this an easy review.