[Feature Request] Add LSTM/GRU Work-Area Size Helper APIs (Caller Preallocation)

[k8ika0s]

Proposed PR branch: qe/issue-08-work-area-size-apis

Category: Feature Request (performance + usability)

Problem

• Callers need deterministic work-area sizing for LSTM/GRU to preallocate and
  avoid runtime allocator overhead in hot loops.

Justification / why this is needed

• Repeated allocation/free in inference loops increases tail latency and
  wastes CPU time under load.
• A sizing helper enables a clean, deterministic preallocation flow for
  frameworks (llama.cpp style loops, ORT EPs, etc.).
• This is additive/opt-in: callers can keep current behavior.

Proposed change

• Add zdnn_get_lstm_work_area_size(...) and zdnn_get_gru_work_area_size(...).
• Add tests.
• (Optional follow-up) Add README guidance/examples for recommended caller preallocation.

Acceptance criteria

• Helper APIs return required bytes (return 0 on invalid inputs).
• Tests cover representative shapes and invalid cases.

Test plan

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

References

• Key files: zdnn/work_area.c, zdnn/zdnn.h, zdnn/zdnn.map,
  tests/testDriver_work_area_size_apis.c.