Fix value-kind mismatch in ffc_negative_digit_comp (float slow path)

[fcostaoliveira]

Summary

ffc_negative_digit_comp (the negative-exponent big-integer slow path) built its rounded-down reference value b with a hardcoded FFC_VALUE_KIND_DOUBLE, then read it back with the caller's kind:

ffc_am_to_float(false, am_b, &b, FFC_VALUE_KIND_DOUBLE);        // writes b.d (8 bytes)
ffc_adjusted_mantissa theor = ffc_to_extended_halfway(b, vk);  // reads b.f (4 bytes) when vk==FLOAT

For vk == FFC_VALUE_KIND_FLOAT this writes the 8-byte double member of the ffc_value union and reads back the 4-byte float member, so theor is garbage and the big-integer tie-break can round the wrong way. The fix is one token: pass the caller's vk.

Impact

Latent for normal inputs (the shortest-form exhaustive sweep never reaches this path). But real for over-precise, hard-to-round float inputs with a negative exponent: a sweep of binary32 midpoints found ~1.07e9 inputs that round to the wrong neighbor under the bug, e.g.

1.1755301539980510070448986385130987351758e-38
  strtof -> 1.17553008e-38 (0x008000ff)   correct
  ffc    -> 1.17553022e-38 (0x00800100)   before this fix (off by 1 ULP)

Tests

• Four midpoint cases added to test_src/float_cases.csv (the harness compares ffc_from_chars<float> vs strtof bit-for-bit). They fail before this change ("Test failures from csvs: 4"), pass after.
• Full float exhaustive (test_exhaustive, all 2^32 values) is "all ok" with the fix; the strtof-parity midpoint sweep drops from ~1.07e9 mismatches to 0. make test passes.