fix: release host weight staging buffer after device upload (47% memory saving)

[masc-it]

Summary

Release the host-side GGUF weight context after all weights have been uploaded to a non-CPU backend.

Previously, device backends such as Metal retained both:

• the complete host-side weight allocation
• the complete device backend allocation

The host context is only needed as a staging source during upload. Once the tensor map points to the device-resident tensors, it can safely be released.

The CPU path is unchanged and continues to use the host context as a zero-copy backend buffer.

Memory improvement

Tested on macOS with Metal using parakeet-tdt-0.6b-v3-q5_k.gguf:

Version	Physical footprint after inference
v0.4.0	1,474 MB
Patched	780 MB
Reduction	694 MB (~47%)

After five consecutive transcription requests, memory remained stable at approximately 771 MB.

The transcription output was identical before and after the change.

The temporary startup peak remains approximately 1.4 GB because both allocations coexist during upload, but the host staging allocation is released immediately afterward.

Validation

• Metal release build succeeds
• Server health endpoint remains healthy
• Five consecutive transcription requests return identical output
• CPU backend produces the same transcript
• Existing non-model test suite passes

Future improvements

This was a no-brainer, zero cost improvement.

For M-series machines, we could approach model loading using two superior approaches:

• Metal mapped buffer ggml_backend_dev_buffer_from_host_ptr
• mmap the GGUF

AI Support

• gpt-5.6-sol

[localai-org-maint-bot]

The device path now releases only the host staging context after all synchronous tensor uploads and after the tensor map has been repointed; the CPU zero-copy path remains unchanged. The ownership/destructor ordering looks sound. I could not rerun the CMake suite locally because this host currently lacks cmake, so my review relies on the focused Metal/CPU validation reported in the PR.

@mudler this looks good to merge.