fix(estimation): let underflowed particles recover

[fzlzjerry]

Fixes #205

What & why

When normalization rounds an individual particle's weight to zero, the next update currently feeds ln(0) into its log-score and makes that particle permanently impossible. Use T::MIN_POSITIVE only for an exactly zero prior weight, preserving the existing numerical path for every representable weight while allowing a later observation to revive an underflowed particle. The public-API regression drives one particle to a displayed zero, reverses the evidence without resampling, and verifies that it becomes dominant.

Checklist

• cargo test + cargo clippy --all-targets clean locally
• New public APIs have a doc example (no new public API)
• No unwrap/expect/panic on library paths (typed errors instead)

Validation

• cargo test -p multicalc --features alloc estimation::particle_filter
• cargo test -p multicalc
• cargo test -p multicalc --features alloc
• cargo test -p multicalc --all-features --doc
• cargo clippy -p multicalc --all-targets --features alloc -- -D warnings
• cargo fmt --all -- --check
• RUSTDOCFLAGS='-D warnings' cargo doc -p multicalc --no-deps --all-features
• git diff --check

[github-actions]

Thanks for the first PR! 🎉 You'll get a first review within 48 h. CI runs the 5-target matrix — a budget-gate failure isn't your fault, see ci/README.md.