mps: amax drops NaN and answers -FLT_MAX for all -inf; randn fails at an F64 -> F32 cast #49

Found by #29's design round (measured on the current build, this Mac):

• aten.amax.default on mps returns -FLT_MAX (f16: -65504, bf16: -3.39e38) for an all--inf input, and drops NaN: amax([nan, 1]) is 1.0. Cause: candle's Metal reduce fold starts from the finite lowest value, not -inf, and skips NaN. The same helper feeds softmax's row max.
• torch.randn(..., device='mps') fails with aten.normal_: Metal contiguous to_dtype F64 F32. eq and to(bool) on an empty mps tensor take an F32 -> F64 path.
• min([0., -0., 0., -0.]) on cpu returns 0.0 where upstream returns -0.0. The golden comparator cannot see the sign of zero.

Done when

• amax/amin/max/min on mps agree with upstream for all-±inf, NaN and ties, using mps: generate stops at zero-element tensors and at the host-readback ops (isin, argmax, ...) #29's first_extremum_index + gather approach. Counters show Metal computed them.
• randn/normal_ on mps compute on the device and agree with upstream's distribution checks (or draws, if the stream can match). Otherwise they refuse by name.
• The golden comparator distinguishes ±0.0, and the min case is fixed.
• Nullified. Gate green.