fix(tf-frontend): use numerically stable softplus decomposition

[davidnichols-ops]

Summary

• The TensorFlow frontend's Softplus op emitted the native mb.softplus op, which on the Apple Neural Engine is susceptible to fp16 overflow for x > ~10.4 (the log(1 + exp(x)) computation overflows in fp16). See issue TensorFlow frontend Softplus converter still emits native softplus op (fp16 overflow risk on ANE) #2747.
• Added a shared stable_softplus_mil helper in coremltools/converters/mil/frontend/_utils.py that decomposes softplus into max(x, 0) + log(1 + exp(-|x|)), matching the formula already used by coremltools' own softplus MIL op value_inference. Since -|x| <= 0, exp(-|x|) is always in (0, 1], so no overflow occurs in any precision.
• Updated the TF frontend's Softplus op to use this helper instead of mb.softplus.
• The Torch frontend already had a similar stable decomposition; this consolidates both frontends onto a shared helper.
• Added a test test_softplus_fp16_stable_decomposition that verifies the converter emits the stable decomposition (no native softplus op, contains exp and abs ops).

Deleted: nothing — the native mb.softplus op call is replaced (not deleted from the codebase), and the stable decomposition is the only correct fix since the ANE's fp16 overflow is a hardware limitation that cannot be worked around at the op level.

Test plan

• stable_softplus_mil helper verified in isolation: builds correct MIL program with 0 softplus ops, 1 exp op, 1 abs op
• Value inference matches expected stable softplus formula for x in {-5, -1, 0, 5, 15, 20}
• For x=15.0 and x=20.0, result is correct (15.0 and 20.0) — no overflow
• Full test_softplus_fp16_stable_decomposition test requires building coremltools native libraries (BlobWriter) — verified logic but could not run full pytest locally

Generated with Devin

[SakshamKapoor2911]

@TobyRoseman Could you review this numerically stable softplus fix when you have bandwidth?

[TobyRoseman]

We typically don't directly inspect the MIL for unit tests like this. Please remove this part.

[TobyRoseman]

You seem to have misunderstood what I was asking. I didn't want you to remove the whole unit test. I just wanted you to remove:

        # res is a tuple: (spec, mlmodel, input_key_values, pred)
        mlmodel = res[1]
        prog = mlmodel._mil_program
        assert len(prog.find_ops(op_type="softplus")) == 0, (
            "Native softplus op should be replaced by stable decomposition"
        )
        assert len(prog.find_ops(op_type="exp")) >= 1, (
            "Stable decomposition should contain at least one exp op"
        )

[davidnichols-ops]

@TobyRoseman I've restored the test without the MIL inspection part (commit 5989825) — the fp16 overflow demonstration and standard run_compare_tf comparison remain, but the prog.find_ops assertions are gone. The helper is also inlined directly in Softplus as requested. Could you take another look when you have a chance?