opencog_ml: proper sampling, double-free fix, warning cleanup, CI inference coverage

[Copilot]

sample_token was greedy-only (temperature/top-k/top-p parameters silently ignored), two inference functions had a double-free in their error paths, and opencog_ml_tests ran without a model so all inference tests were skipped in CI.

Sampling (opencog_ml.cpp)

Replaced stub greedy implementation with full pipeline:

• Softmax with max-subtraction for numerical stability
• Top-k: zero out all but highest-k logits
• Nucleus (top-p): sort by probability, zero out below cumulative threshold
• Temperature scaling: raise probs to 1/temperature, renormalize
• Multinomial draw from surviving candidates

Greedy path preserved when temperature <= 0.

Double-free fix (opencog_ml.cpp)

In both opencog_ml_infer_token and opencog_ml_infer_sequence, owns_data = true was set on the tensor views before the NULL check. If either create_tensor_view call failed, the error path called both free(buffer) and opencog_ml_free_tensor() on the same pointer. Fixed by moving owns_data = true to after the check, keeping raw-buffer frees only in the error branch.

Compiler warnings (test_opencog_ml.c, opencog_ml_example.c)

• Added void to C function prototypes (-Wstrict-prototypes)
• Removed has_valid_data set-but-not-used variable
• Cast float → double in printf calls (-Wdouble-promotion)

CI inference coverage (CMakeLists.txt)

Passed tiny-rwkv-4v0-660K-FP32.bin to opencog_ml_tests so engine-management, state-management, and basic-inference test cases actually execute in CI instead of printing SKIP.