Add polynomial-interpolation cut-cell stencils (E2); harden devcontainer build

[pbrady]

Summary

Migrates the project's Mathematica polynomial-interpolation cut-cell derivation into the SymPy stencil framework (scripts/stencil_gen/), and hardens the devcontainer's Spack build.

Poly-interpolation stencils (the new capability)

Define interpolation polynomials f[i+y] = Σ_j γ_ij(y, ψ, params)·f[node_j] (TEMO-style); the derivative stencil then falls out as (1/h)·d/dy γ|_{y=0}. This yields one consistent free-parameter set — fa feeds both interp_wall and nbs_floating; ia is interpolation-only; da is the Dirichlet derivative.

• stencil_gen/interp.py (new): derive_interior_interp, derivative_from_interp (the d/dy|0 core), cut-cell wall rows (two-variant zero+blend+average), derive_floating_coeffs via the interp↔derivative duality, and a derived Dirichlet closure (poly recipe + one conservation solve). The transcribed dirichlet_coeffs_symbolic is retained as a cross-check oracle.
• taylor_system.py: _interp_rhs / build_interp_system (eval-at-y functional).
• printer.py, codegen.py: emit real interp_interior / interp_wall / query_interp. Additive and gated on None-default StencilGenSpec fields, so existing stencils emit unchanged. __main__.py: generate polyE2_1.
• tools/eval_poly.py: parse a polyE2_1.cpp's runtime methods for numeric comparison.
• Tests (tests/test_interp.py, tests/test_codegen_poly.py): symbolic golden values, interp↔derivative duality, polynomial exactness, value/derivative decoupling, derived-vs-oracle Dirichlet, and a capstone asserting the regenerated polyE2_1.cpp is numerically equivalent (≤1e-12, all four runtime methods) to the committed src/stencils/polyE2_1.cpp.
• docs/poly_interp_design.md, _proto_interp.py: design spec + a runnable validation demo.

Pure-Python and additive — src/stencils/polyE2_1.cpp is unchanged (the generator + capstone prove the codegen stays compatible). Scope is E2 only for now.

Devcontainer

• .devcontainer/Dockerfile: spack install --fail-fast. The default best-effort install skips failed packages and their dependents yet exits 0 — which could produce a dependency-less image that still looked successfully built. --fail-fast aborts on the first failure. Pairs with the SPACK_MIRROR_URL build-arg for reliable source fetches behind a TLS-intercepting proxy.

Validation

• t-polyE2_1 passes in a real C++ build (verified both in the CI dependency image and a native arm64 devcontainer build).
• The capstone test confirms the regenerated codegen output matches the committed C++ to 1e-12 across all four runtime methods.
• The new SymPy suite passes; existing stencil_gen tests are unaffected (additive, gated behind None-default fields).

Follow-ups (not in this PR)

• Generalize beyond E2 (the Dirichlet closure's conservation step is currently the E2 specialization).
• Optionally regenerate src/stencils/polyE2_1.cpp from the pipeline (numerically equivalent output; left as-is to avoid a noisy CSE diff).