[rebranch] Adjust to "[AsmWriter] Change the output syntax of floating-point literals. (#190649)"
```
ConstValues/SectionIR.swift:64:
expected: // CHECK: @"$s9SectionIR13floatLiteral1Sfvp" = {{.*}}constant %TSf <{ float 0x40091EB860000000 }>, section "mysection"
actual: @"$s9SectionIR13floatLiteral1Sfvp" = hidden constant %TSf <{ float 3.140000e+00 }>, section "mysection", align 4
```
The offending commit reworked `writeAPFloatInternal` so that the legacy
hexadecimal literal form is never emitted. A decimal literal is now used
whenever reparsing the 6-digit decimal string *in the value's own
semantics* yields a bit-identical `APFloat`; only otherwise does the
printer fall back to the new exact `f0x...` form. Previously `float` was
widened to `double` and printed as a 64-bit hex pattern.
`float 3.140000e+00` parses to 0x4048F5C3, which widens to exactly
0x40091EB860000000, so the new spelling denotes the very same bits: the
change is purely cosmetic, not a loss of precision. Re-pinning the exact
decimal spelling matches how these tests already spell the neighboring
`double 2.718000e+00` and integer constants, and mirrors the downstream
adaptation of a clang test in llvm-project b91eb068c00c.
See https://github.com/llvm/llvm-project/commit/41c214f0b1151d74b660e5218b9ef066f4207492 (https://github.com/llvm/llvm-project/pull/190649).
Assisted-by: Claude Code
Cherry-picked to main, which is still paired with llvm-project
stable/21.x. The old AsmWriter there still prints the widened 64-bit hex
pattern, so re-pinning these four CHECK lines to the decimal spelling
alone would fail on main while passing on rebranch. Match either spelling
with a FileCheck alternation instead, so one test file works against both
LLVM versions. Both alternatives denote the same bits (float
3.140000e+00 parses to 0x4048F5C3, which widens to exactly
0x40091EB860000000), so nothing is loosened beyond the spelling. Collapse
this to the decimal form alone when main moves to LLVM 23.