FPU variants: math.pow(-1.0, NaN) returns 1.0 instead of NaN (__wrap_pow not applied on 68020fpu/68040)

[sidick]

Summary

On the FPU variants (68020fpu, 68040), math.pow(-1.0, float("nan"))
returns 1.0 instead of NaN, so the NaN-propagation guard fails:

>>> import math
>>> math.pow(-1.0, float("nan"))
1.0          # expected: nan

This trips tests/ports/amiga/test_floatconv_smoke.py:57
(assert math.isnan(math.pow(-1.0, nan))) with an AssertionError.

Affected / not affected

Variant	pow(-1, NaN)	floatconv test
standard (soft-float)	nan ✅	passes
68020fpu	1.0 ❌	fails @ line 57
68040	1.0 ❌	fails @ line 57

Root cause (hypothesis)

The port works around libnix/bebbo's non-IEEE pow(-1, NaN) -> 1.0 with a
linker-wrapped __wrap_pow (ports/amiga/floatconv.c:260, wired via
-Wl,--wrap=pow,--wrap=tgamma in ports/amiga/Makefile:97).

The wrap is effective on the soft-float build but appears not to be
applied on the FPU builds — the FPU variants presumably reach pow
through a different symbol/inlined FPU path that --wrap=pow doesn't
intercept, so the unpatched libnix behaviour leaks through. The
tgamma wrap (gamma(-inf)) likely has the same exposure and should be
checked alongside.

Reproduction

ports/amiga/tools/amiga-build.sh 68020fpu
# On Amiberry (68020/68881-82) or real FPU hardware:
micropython tests/ports/amiga/test_floatconv_smoke.py
# -> Traceback ... line 57, in <module>  AssertionError

Cannot be reproduced under vamos — its 68040 FPU emulation is not faithful
enough (the test is documented standard-only for vamos for that reason).
Verified on Amiberry with cpu_type=68020/68881, fpu_model=68882.

Pre-existing

Not a regression in the v1.28.0-amiga release — the current development
branch (amiga-port) FPU binary fails identically at the same line on
Amiberry. Filing so the FPU-variant __wrap_pow/__wrap_tgamma gap is
tracked independently of the release.

Possible directions

• Confirm whether pow/tgamma are actually emitted as calls on the FPU
  builds (check the linker map / nm for __wrap_pow vs pow references).
• If the FPU build inlines or routes math through a different symbol,
  extend the wrap set or handle NaN propagation in floatconv.c in a way
  that survives the FPU codegen path.
• Add an FPU-variant assertion to the test matrix once fixed.