BUG: AVX-512F log() overflows

LLVM has a bug with AVX-512F masked multiply intrinsic `_mm512_mask_mul_pd`.  The multiply operation is performed unmasked with a masked store.  This can cause FP exceptions for lanes that should be masked.

LLVM bug report: https://bugs.llvm.org/show_bug.cgi?id=51988.

The bug affects NumPy / SciPy because NumPy has an AVX-512F implementation for log(), which uses a masked multiply intrinsic: `_mm512_mask_mul_pd`.

The bug surfaces in SciPy when trying to setup constants:
```
# The largest [in magnitude] usable floating value.
_XMAX = np.finfo(float).max

# The log of the largest usable floating value; useful for knowing
# when exp(something) will overflow
_LOGXMAX = np.log(_XMAX)
```

The workaround here explicitly masks the inputs before the multiply operation.

**Testing**
- Added a test case to numpy/core/tests/test_umath.py
  - Fails on trunk
  - Passes after workaround

SciPy testing before:
```
93 failed, 22378 passed, 1385 skipped, 93 xfailed, 7 xpassed, 1128 errors in 117.54s (0:01:57)
```

SciPy testing after:
```
74 failed, 33361 passed, 2230 skipped, 108 xfailed, 8 xpassed in 178.19s (0:02:58)
```

Note, all `1128` errors have been resolved as well as less test failures (`93` --> `74`)

0 files changed, 0 insertions, 0 deletions