Add M nearest-neighbour Chatterjee correlation (#990)

[su-senka]

Summary

Implements the revised (M nearest-neighbour) Chatterjee rank correlation of
Lin & Han (2021), addressing #990. Adds a new function
chatterjee_correlation_mnn(u, v, M) alongside the existing
chatterjee_correlation, with the same C++11 and C++17 overload structure.

The original coefficient has a detection boundary of n^(-1/4) for independence
testing, well short of the parametric n^(-1/2) rate. By using the M right
nearest neighbours of each point (rather than the single right neighbour) and
letting M grow with n, the revised statistic consistently estimates the same
dependence measure while approaching near-parametric efficiency. See Lin & Han,
On boosting the power of Chatterjee's rank correlation, Biometrika 110(2)
(2023) 283–299, arXiv:2108.06828.

Design notes

• Separate function rather than an extended signature. The M-NN statistic
  uses min(R_i, R_j) and a different normalisation, so even at M = 1 it is not
  identical to chatterjee_correlation. A distinct function avoids silently
  changing existing results and keeps the statistical intent explicit.
  M is a required argument with no default.

• Rank base. The internal rank() returns 0-based ranks; the paper's
  formula uses 1-based ranks. The offset cancels in the existing M = 1 statistic
  (which uses |R_i - R_{i+1}|) but not under min(.,.), so it is applied
  explicitly. This is noted in a comment where it matters.

• Complexity. O(n log n + nM). Near-linear for small M; tends to O(n²) as
  M → n.

• Parallel path. The outer index loop is partitioned across threads into
  disjoint ranges, each reading the shared rank vector read-only (indices up to
  i + M may fall in a neighbouring range; there are no writes). This differs
  from the M = 1 parallel path, which splits the data array for the
  difference-based transform.

• Ties / degenerate input. Like chatterjee_correlation, the function
  assumes distinct Y (continuous data). A constant Y returns a quiet NaN; this
  is detected on the input directly, since rank() collapses tied values.

• Choice of M. The asymptotic null variance is minimised at M ~ sqrt(n); the
  choice is documented but left to the caller.

Tests

Added to test_chatterjee_correlation.cpp, covering float, double, and
long double:

• Exact closed-form checks against the paper's Remark 2.5 (strictly increasing
  and strictly decreasing dependence), which require no external reference.
• Small exact spot values computed independently as rationals.
• Constant-Y → NaN, and invariance under strictly increasing transforms of X
  and Y.
• Sequential/parallel agreement across several M (under the parallel build).

The sequential path was verified locally under b2 with cxxstd=14 and
cxxstd=17 (clang, arm64).