Add tests and benchmarks for truncated eigendecomposition in p-value computation

[Kinara2020]

Following up on #550, which replaced the pseudoinverse-based p-value
computation with a rank-r truncated eigendecomposition per Wood (2013).

This issue tracks remaining work to validate and document that change:

Testing gaps:

• Edge case: rank-deficient covariance matrices
• Edge case: near-zero eigenvalue threshold sensitivity
• Numerical comparison of p-values between Wood (2006) vs Wood (2013)
  methods on known datasets

Benchmarking:

• Runtime comparison: truncated vs full eigendecomposition across
  varying smooth term sizes

References:
Wood, S.N. (2013). "On p-values for smooth components of an extended
generalized additive model." Biometrika 100(1):221–228.