Check index input coordinates in one pass #217

Problem

Since 0.8.0 (#195), indexes reject non-finite and out-of-range coordinates with
check_coordinates in include/rabitqlib/utils/input_checks.hpp. It reads the input twice
in serial: all_finite, then a range check. On 1M × 960 float32 vectors the check alone took
about 0.70 s, which accounts for most of the 0.6–0.7 s that IVF construction gained between
0.7.1 and 0.8.0 in our benchmark.

The range check !(std::fabs(v) <= bound) is already true for NaN and ±Inf, so one pass can
detect every bad value; the finite check is only needed to choose the error message.

Task

Make check_coordinates read the data once in the common (valid) case:

1. Count out-of-range values in one pass.
2. Only if that count is non-zero, call all_finite to choose between the two existing
   messages.

Done when

• Error types and messages are unchanged, including which one is raised when the input has
  both NaN and large finite values (non-finite wins today).
• Existing input tests pass: tests/python/test_input_boundaries.py and the C++ unit tests.
• The PR reports the check's time before and after on a large array (for example 1M × 960
  random float32), with the CPU, compiler and build type.

Notes

• This header is included by ISA-specific translation units (src/fastscan/estimator_avx2.cpp
  and src/fastscan/estimator_avx512.cpp). Besides
  ./scripts/check-format.sh and the C++ tests, run ./scripts/check-isa-symbols.sh on a
  build with RABITQ_ENABLE_NATIVE_OPTIMIZATION=OFF (see
  verification by change type).
• Parallelizing the scan is a possible follow-up; keep this PR to the single pass.

Setup: contribution guide, C++ changes.