docs: correct overstated claims, cite the Go issues, state measured results

[tannevaled]

A review of the round-2 write-ups against the measurements:

• README: the SIMD kernels exist on amd64 and arm64 only (it said all six
  targets, selected at run time); it now lists them all (not only the
  sum), and the performance summary states what was measured on a Zen 3
  against NumPy 2.5.3, including where go-ndarray is slower, instead of
  only the Apple-silicon parity results.
• perf.md: the parallel-GEMM stall and NumPy's threaded-dot advantage are
  hypotheses, not findings (the GEMM drop already shows at 2 cores of one
  core complex, which the shared-panel explanation does not cover); the
  "40% in the edge loop" profile covered 64^2 and 512^2 together; the
  Workspace table is best of two, not median; 43 GFLOP/s is not a
  measured peak; ULP figures name their sample sizes.
• The amd64 math.Exp bug is now math: Exp returns +Inf for finite results on amd64 (709.436 <= x < 709.783) golang/go#81995 (reproduced with Go 1.26.8
  and 1.27.1 on Zen 3, Cascade Lake and Haswell, with and without FMA);
  the math.Log one was already math: math.Log does not handle subnormal floating point number on amd64 golang/go#56600. The note that math.Exp
  returns 0 just above ln(2^-1075) is about arm64, not amd64, and is a
  last-place difference, not a bug.

The README example was compile-checked. No code change.

🤖 Generated with Claude Code