[Feature Request] Add Prepared Quantized Matmul Context APIs (Prepare/Execute/Release)

[k8ika0s]

Proposed PR branch: qe/issue-12-prepared-quantized-matmul-context

Category: Feature Request (performance for repeated-shape workloads)

Problem

• Repeated quantized matmul calls with stable descriptors redo setup each time.
• A prepared context reduces per-call overhead for static-shape workloads.

Justification / why this is needed

• Inference workloads often reuse stable tensor descriptors (weights fixed,
  shapes repeated). Paying setup costs per call is avoidable overhead.
• Prepared contexts let frameworks precompute/validate once and reuse safely.

Proposed change

• Add:
  • zdnn_prepare_quantized_matmul_context(...)
  • zdnn_quantized_matmul_op_prepared(...)
  • zdnn_release_quantized_matmul_context(...)
• Keep prepared-context work-area management self-contained (internal aligned
  buffer management in the context lifecycle).
• Add tests.
• (Optional follow-up) Add README guidance for the prepared-context lifecycle.

Acceptance criteria

• Prepared path matches non-prepared behavior (status + results).
• Context lifetime rules are enforced (no leaks, safe reuse).
• Tests cover prepare/execute/release + invalid lifecycle sequences.

Test plan

• make test (expects tests/testDriver_quantized_matmul_prepared_context.c to pass).

References

• Key files: zdnn/prepared_quantized_matmul.c, zdnn/zdnn.h, zdnn/zdnn.map,
  tests/testDriver_quantized_matmul_prepared_context.c.