Fix #23802 - don't fuse conversion ops on common subexpressions

[limepoutine]

If a floating point conversion is a common subexpression, avoid encoding precision in the ModR/M byte, but instead do a full load. Otherwise global CSE will have a hard time finding the value in registers.

@WalterBright Please review as this might not be an optimal fix.

[github-actions]

DMD perf check

Metric	Base	PR	Δ
compile Phobos codegen (instr)	1,472.4 M	1,475.4 M	+0.200%
dmd binary size (stripped)	6.87 MB	6.88 MB	+0.11%

All measurements

Metric	Base	PR	Δ
compile hello.d (instr)	213.3 M	213.3 M	+0.007%
compile hello.d -O -release (instr)	231.6 M	231.6 M	+0.010%
compile Phobos (instr)	5,120.6 M	5,123.4 M	+0.056%
compile Phobos codegen (instr)	1,472.4 M	1,475.4 M	+0.200%
compile vibe.d (instr)	15,112.0 M	15,112.3 M	+0.002%
dmd binary size (stripped)	6.87 MB	6.88 MB	+0.11%
hello binary size (stripped)	0.72 MB	0.72 MB	0.00%
peak RSS (compile hello.d)	43.16 MB	43.19 MB	+0.07%
peak RSS (compile Phobos)	620.4 MB	617.2 MB	-0.51%
peak RSS (compile vibe.d)	1918 MB	1917 MB	-0.06%
compile dmd itself (wall)	12.0 s	12.1 s	+0.37%
compile hello.d (wall)	62.7 ms	62.7 ms	+0.08%
compile Phobos (wall)	1,544 ms	1,548 ms	+0.24%

_{c97396d vs merge-base 74623e4 · about these metrics}