Optimizations made:
- `extract_coefficients` (util.jl): Fixed type instability with `findfirst`
  returning Union{Nothing, Int}. Now uses explicit type-stable loop.
  Also pre-computes array lengths outside loops for better performance.

- `monomial_compress` (wronskian.jl): Replaced `Array{Any, 1}` with typed
  `Vector{Tuple{P, T}}` for type stability. Pre-computes parameter names
  in a Set for O(1) lookup instead of repeated map operations.

- `massive_eval` (wronskian.jl): Uses typed containers (Set{Vector{Int}},
  Dict{Vector{Int}, T}) instead of untyped. Pre-allocates working arrays
  and uses in-place operations with @inbounds. Pre-sizes result array.

- `det_minor_expansion_inner` (elimination.jl): Replaced `in keys(cache)`
  with `haskey(cache)` for better performance. Pre-allocates Sets for
  discarded rows/cols. Uses `sort!` on mutable arrays instead of allocating.

These changes improve type stability and reduce unnecessary allocations
in hot code paths, particularly benefiting larger ODE systems.

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

Per reviewer feedback, reverted complex performance-focused changes
that add complexity without making the code cleaner. Kept only
changes that genuinely improve code quality:

Kept:
- haskey() instead of `in keys()` (standard Julia idiom)
- sign = -sign (cleaner)
- Typed arrays Vector{Tuple{P,T}} instead of Array{Any,1}
- Pre-computed param_names Set (avoids repeated map)
- Typed Set{Vector{Int}} and Dict in massive_eval
- Removed commented-out dead code

Reverted:
- Complex pre-allocated loops in det_minor_expansion_inner
- Explicit loops replacing map() in extract_coefficients
- Pre-allocated working arrays and @inbounds loops in massive_eval

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>