[ajnonaka]

Adds an unsplit convolutional PML (Roden & Gedney) for the Yee FDTD solver, selected with warpx.pml_type = convolutional. The default remains split.

Also fixes a bug in how the PEC mask (algo.use_PEC_mask) is applied, which affects split-field runs too (see below).

CPML
-CPML E, B and H (LLG) updates in the existing PML boxes, using the same σ profile as the split-field PML (warpx.pml_ncell, warpx.pml_delta). The total field is stored in component 0, so the existing exchange, fill-boundary, PEC and checkpoint code is reused.
-Works with spatially varying ε, μ and σ, using the same Lax-Wendroff or backward-Euler material update as the interior.
-Supports the PEC mask and its extension into the PML, including a second layer (sigma_npy_file2, npy_k_index2).
Optional stretching parameters warpx.cpml_kappa_max and warpx.cpml_alpha_max, plus their grading orders.
-The CPML auxiliary variables are written to checkpoints, so restarts reproduce an uninterrupted run exactly.
Aborts for unsupported combinations: CKC, nodal or PSATD solvers, divergence cleaning in the PML, J-damping, moving window.

PEC-mask fix (changes results of existing use_PEC_mask runs)
-Before: the main-grid E was masked only after the guard-cell/PML exchange. The exchange therefore sent unmasked tangential E on PEC edges to the PML and to neighbouring grid boxes. The next B update saw inconsistent E across the PML interface along the PEC layer. This caused an instability that showed up as NaNs at the PEC layer next to the PML in long runs.
-Now: the main-grid E is masked right after its update, before the exchange. It is still masked again after the excitation, so masked edges stay exactly zero.
-This applies to both PML types, so results of existing PEC-mask runs will change slightly.

New inputs
warpx.pml_type: split (default) or convolutional
warpx.cpml_kappa_max (default 1), warpx.cpml_alpha_max (1/s, default 0), warpx.cpml_grading_order (default 2), warpx.cpml_alpha_grading_order (default 1)
algo.extend_PEC_mask_into_pml (default 1, same behaviour as before; 0 = no PEC mask in the PML)

All are documented in parameters.rst.

Testing (3D, GCC, CPU)
-Reflection compared against a large-domain reference for vacuum, uniform ε_r = 11.7, a layered substrate, and a microstrip defined by a PEC mask. With default settings the CPML matches the split-field PML in each case. Tuned settings (pml_delta = 17 plus small κ/α) reduce reflection by up to about 3×.
-PEC strip running into a lo-x CPML (5.5 GHz soft source):
--before the fix, the PML field blew up within 30k (tuned) to 180k (default) steps;
--after the fix, it stayed stable to 500k steps (tuned) and 210k steps (default, end of run).
-Two PEC layers: stable for 60k steps.
-Restart: checkpoint/restart is bit-identical.
-Other builds: LLG and 2D XZ syntax-checked only, not run. GPU and Windows builds not tested locally.