docs: gpu_solver_architecture.md 'Current State' section says lax.scan and spsolve, both now incorrect

[github-actions]

Problem

docs/gpu_solver_architecture.md contains a "Current State" section (around line 283) that lists major improvements since the document was written:

## Current State

Since this document was written, major improvements have been made:
1. Adaptive timestepping: BDF2 with LTE control is implemented
2. Unified device models: All devices use OpenVAF-compiled Verilog-A (PSP103, etc.)
3. GPU-resident time loops: Transient uses `lax.scan` for full GPU execution
4. Sparse solver: BCOO/BCSR with `spsolve` for large circuits
5. AC and noise analysis: Frequency-domain analyses are available

Two statements in this list are now inaccurate:

1. Line 3: "Transient uses lax.scan for full GPU execution" — the production transient loop (full_mna.py) uses lax.while_loop, not lax.scan. lax.scan requires a fixed number of steps upfront, which is incompatible with the adaptive timestepper.

2. Line 4: "Sparse solver: BCOO/BCSR with spsolve" — the production sparse solver uses Spineax/cuDSS on CUDA GPUs and UMFPACK via FFI on CPU. spsolve (JAX's experimental cuSOLVER wrapper) is not used in the hot path.

Note: issue #98 already covers the historical banner's reference to analysis/dc.py. This issue is about the separate "Current State" section which is supposed to describe the present implementation.