erts: Fix float16 construction in the software fallback

[a1x-an]

Fixes GH-11332.

The software float16 conversion used when _Float16 is not available
(notably all official Windows/MSVC builds) produced wrong bit patterns for
nearly all values that are subnormal in binary16, and misrounded some normal
values, so the emulator could not re-construct a value it had just matched:

1> <<F:16/float>> = <<0,51>>.   %% 51 × 2⁻²⁴, exactly representable
2> <<F:16/float>>.
<<0,39>>                        %% construct(match(B)) ≠ B

Three defects in f32_to_f16() (erts/emulator/beam/erl_bits.c):

1. The subnormal branch OR:ed the implicit leading bit into position 23
   after the mantissa had already been reduced to its top 10 bits, then
   shifted by an exponent-derived amount - producing garbage for all
   binary16 subnormals except a few accidental fixed points. The rounding
   increment was also a full ULP (1 << shift) where half an ULP is
   required.
2. The normal branch ignored bits 9..0 in its sticky test, so values just
   above a rounding midpoint whose sticky bits lie only in the low bits
   rounded down as if they were ties (when the kept LSB was even).
3. Converting through float (f32_to_f16((float) x)) rounds twice
   (double→float→binary16), so results could differ from the native
   single-rounding (_Float16) x path even with 1) and 2) fixed:
   <<1.000488282181322574615478515625:16/float>> (= 1 + 2⁻¹¹ + 2⁻³⁰)
   gave 0x3C00 through the fallback but 0x3C01 on native builds.

The fallback is replaced by a direct integer-exact double→binary16
conversion with a single round-to-nearest-even step, giving results
identical to the native path. The match direction (f16_to_f32) was
correct and is unchanged.

Verification

• Standalone harness against a compiler-provided _Float16 oracle
  (gcc 13): exhaustive sweep of all finite float32 inputs - 0 mismatches for
  the new conversion (the current one disagrees on 417,333,236 inputs);
  construct∘match identity over all finite binary16 values - 0 failures
  (currently 2026); 10⁹ random finite doubles - 0 mismatches.
• Cross-compiler check: the same source compiled with MSVC (VS 2022,
  cl /O2 - the toolchain of the official Windows builds, which lacks
  _Float16 and therefore uses this fallback) and with gcc -O2
  produces bit-identical outputs for both the old and the new conversion
  over all finite float32 inputs (FNV-1a digests over the full sweep
  match exactly), so the fix is compiler-independent.
• Patched emulator built from maint on x86_64 Linux (musl/gcc):
  • native _Float16 path: all witness values and the full binary16
    identity sweep pass (unchanged behaviour);
  • fallback path forced by undefining FLOAT16_IS_CONVERTIBLE: all
    witness values and the full identity sweep pass, including the
    double-rounding witness - the fallback now matches the native path
    exactly.
• The shipped Windows 29.0.3 emulator (which uses the fallback)
  reproduces all failures; the new bs_construct_SUITE:fp16/1 cases fail
  on it and pass with the fix.

New test cases cover exact subnormals (previously corrupted), a
sticky-bits rounding witness, an exact tie (round-to-even), the
double-rounding witness, and a construct∘match identity sweep over every
finite binary16 value.

[CLAassistant]

All committers have signed the CLA.

[github-actions]

CT Test Results

    3 files    136 suites   50m 49s ⏱️
1 682 tests 1 626 ✅ 56 💤 0 ❌
2 325 runs  2 251 ✅ 74 💤 0 ❌

Results for commit 6530318.

♻️ This comment has been updated with latest results.

To speed up review, make sure that you have read Contributing to Erlang/OTP and that all checks pass.

See the TESTING and DEVELOPMENT HowTo guides for details about how to run test locally.

Artifacts

• Complete CT logs (Download Logs)
• HTML Documentation (Download HTML Docs)
• Windows Installer

// Erlang/OTP Github Action Bot

[jhogberg]

Thanks for the PR! Can you rebase this onto the tag patch-base-28 so that we can release a patch on OTP 28? :-)

[a1x-an]

Done, rebased onto patch-base-28 :)