fix: update cppdlr symmetric grid dependency

[weiyiguo9]

Summary

This PR integrates the upstream symmetric-DLR mesh fix from flatironinstitute/cppdlr#20, following the divergence reported in flatironinstitute/cppdlr#19 for GW/GF2 calculations.

• Pin cppdlr to commit 6f3f2d836d43da7a01b4e53292ad49f896a19667 from Include self-symmetric fixed points in symmetrized DLR grids flatironinstitute/cppdlr#20 so symmetric DLR grids include the self-symmetric fixed points.
• Reuse an existing c2py::c2py target when NDA/h5 has already provided c2py, avoiding duplicate top-level FetchContent setup.

Build and unit-test validation

• The corresponding Snellius wg-dev worktree configured, built, and installed successfully with the exact cppdlr commit above.
• Focused tests passed: test_iaft_dlr, test_iaft (2/2).
• The installed Python-enabled CoQui build was then used for the material calculations below.

Material validation

All calculations used the PR build with a high-precision DLR basis, beta = 300, Lambda = 10000, and fresh (restart = false) SCGW runs. The PR20 mesh was reported as (nt_f, nt_b, nw_f, nw_b) = (89, 89, 90, 89).

Case	Mesh / bands	Iteration scheme	Result
SrVO3	4x4x4, 150 bands	commutator DIIS, mixing 0.6, warmup 2, subspace 6	Converged at iteration 12: dFock = 4.90e-8, dSigma = 6.85e-8
fcc Al	5x5x5, 200 bands	commutator DIIS, mixing 0.10, warmup 2, subspace 6	Converged at iteration 9: dFock = 1.47e-7, dSigma = 5.46e-7
bcc Fe (FM)	5x5x5, 200 bands, spin-polarized	commutator DIIS, mixing 0.20, warmup 2, subspace 6	Did not fully converge through iteration 11; the bounded oscillations, occasional spike, and subsequent recovery are qualitatively similar to the existing IR behavior

For SrVO3, a separate damped PR20-DLR run without DIIS also followed a smooth IR-like trajectory rather than the old DLR divergence. After 24 iterations it reached dFock = 4.65e-4 and dSigma = 2.84e-3 and remained stable throughout.

The Fe test is intentionally reported as preliminary rather than converged. Its residuals decreased during warmup, oscillated after DIIS extrapolation, showed a spike at iteration 9, and recovered after DIIS history eviction, but did not establish a trend toward the 1e-6 threshold. This resembles the previous difficulty in fully converging the corresponding IR calculation and does not indicate a return of the old symmetric-DLR mesh pathology.

The material tests were also repeated across topology-aware MPI layouts. The first-iteration physical values agreed across legal layouts. SrVO3 correctly rejected an oversized 384-rank processor grid because its 150-band dimension cannot support a grid axis of 192; this is a CoQui distributed-array layout constraint, not a cppdlr failure.

Why this is ready to merge

The focused DLR tests pass, and the same PR build now converges both SrVO3 and Al real-material scGW controls below the 1e-6 threshold. An additional Al GF2 direct-term control converged in 18 iterations with essentially the same residual trajectory and final result as IR, confirming that the PR20 symmetric-grid fix also applies to GF2. The separate damped SrVO3 run also remains stable without DIIS, showing that the result is not dependent on DIIS hiding the old symmetric-grid instability. FM Fe does not fully converge, but its oscillation and recovery are qualitatively similar to the existing IR behavior and do not reproduce the old DLR-specific divergence. Together, these results show that the upstream cppdlr fix addresses the reported regression without requiring an additional CoQui-side DLR workaround. I therefore believe this dependency update is ready to merge; further Fe iteration-algorithm tuning can be handled independently.

Acknowledgement

This bug detection and comparison is also suggested by Xiansheng Cai