refactor: omtimize subspace diag freq in bpcg

[Missing-Hex]

Summary

Optimize BPCG subspace diagonalization frequency with dynamic scaling based on problem size (n_band × n_basis), replacing the fixed frequency (nline = 4).

Problem

Current implementation calls calc_hsub_with_block every fixed nline = 4 CG iterations:

• Small systems: Unnecessary overhead (O(n_band³) cost dominates)
• Large systems: Insufficient orthogonality maintenance, slower convergence

Solution

New function compute_optimal_freq() with thresholds:

Problem Size	Frequency
≥ 10,000,000	2
≥ 2,000,000	3
≥ 500,000	4 (baseline)
≥ 100,000	5
≥ 10,000	6
< 10,000	nline

Code Change

// bpcg_kernel_op.h
int compute_optimal_freq(const int n_band, const int n_basis, const int nline);

// diago_bpcg.cpp - before CG loop
this->optimal_freq = compute_optimal_freq(this->n_band, this->n_basis, this->nline);

// CG loop
if (current_scf_iter == 1 && ntry % this->optimal_freq == 0) {
    this->calc_hsub_with_block(...);
}