Improve intermediate precision in exact 1D clustering

[rioyu123]

Summary

Clustering1D::train_exact() can lose precision in two intermediate calculations:

• The interval-cost prefix sums are stored as double, but x * x is evaluated as float first. The rounded squared values can change the partition selected by the dynamic program when the input has a large common offset.
• Cluster means are accumulated in float. For example, a single cluster of 128 identical 1000001.0f values currently returns 1000000.125f instead of 1000001.0f.

Promote each value before squaring and accumulate the cluster sum in double. Centroids remain float; the SMAWK recurrence, cost-table types, public API, and asymptotic complexity are unchanged. Numerical results are not bit-identical: removing intermediate float rounding can change centroids even on well-conditioned inputs, and can also change the selected partition.

The regression tests use the public Clustering1D::train_exact() path. They cover a known eight-cluster partition before and after exactly representable positive/negative offsets, including float output rounding, and constant-valued clusters. This is a targeted precision fix: the cost table still uses float, and prefix-sum subtraction remains subject to cancellation.

Validation

• Built the CPU-only generic faiss_test target on Linux against main 6bbb068ad8dab612b84304400846da1b14539085.
• Both new tests fail with the unchanged implementation and pass with this patch.
• Full public C++ CTest run: 277 passed, 5 skipped (SIMD-specific checks in this generic build), no failures.
• Additional standalone checks compare small cases against an independent quadratic-time dynamic-programming reference, with an allowance for float centroid rounding.
• Changed C++ lines/new test formatting and git diff --check pass.
• No Python bindings (including tests/test_clustering.py), GPU tests, or Meta-internal tests were run.