[j-olivier]

__JS_NewFloat64() canonicalizes NaN with an integer test on the bit pattern. This replaces it with a
floating point comparison:

-    if (js_unlikely((u.u64 & 0x7fffffffffffffff) > 0x7ff0000000000000))
+    if (js_unlikely(d != d))

Generated code:

target	before	after
x86-64 (gcc -O2)	8 instructions, 2 constants	5 instructions, 1 constant (ucomisd + jp)
arm64 (clang -O2)	9 instructions	7 instructions, branch free (fcmp + csel)

V8 benchmark, 10 interleaved runs each, pinned to two logical processors at high priority (Intel
Cascade Lake, mingw-w64 gcc -O2). NavierStokes improves by 8.2% (median +6.9%), outside the run to
run noise (sd 0.7% to 3.0%); the other benchmarks are within the noise:

	before	sd	after	sd	change
Richards	1275	0.6%	1279	1.1%	+0.3%
DeltaBlue	1152	1.2%	1160	1.2%	+0.7%
Crypto	1473	2.3%	1497	1.0%	+1.7%
RayTrace	2124	2.1%	2085	3.0%	-1.9%
EarleyBoyer	2737	1.7%	2768	0.7%	+1.1%
RegExp	518	1.7%	515	2.5%	-0.7%
Splay	4302	1.8%	4333	1.2%	+0.7%
NavierStokes	2712	3.0%	2933	0.7%	+8.2%
Score	1718	1.5%	1739	0.9%	+1.2%

Verification:

• Old and new __JS_NewFloat64() give identical results on 100M values (random bit patterns, all
  sign/exponent classes, all single-bit NaN payloads, sNaN/qNaN, +-0, +-Inf, denormals), with an exact
  round trip through JS_VALUE_GET_FLOAT64() for non-NaN values. Tested on x86-64 (shift 47, gcc
  -O0/-O2/-O3) and 32-bit x86 (shift 32, clang, SSE2 and x87, -O0/-O2).
• make test and test262 (x86-64): unchanged results.
• arm64: code generation checked, not run.

d != d is not compatible with -ffast-math / -ffinite-math-only; they are not used by the Makefile
and quickjs.c already relies on IEEE NaN semantics (isnan()).