perf: FxHash node-cache + direct attr-value reads (output-identical, ~28-30% on node-heavy docs)

[dginev]

Two dependency-free, output-identical performance improvements to the hot paths
exercised by node/attribute-heavy XML processing (measured downstream in
latexml-oxide, a Rust port of LaTeXML).
Both are based on the latest main.

1. perf(node-cache): FxHash-style hasher for the xmlNodePtr -> Node map

The internal node-bookkeeping map is probed on every Node::wrap/lookup and used
the std SipHash RandomState. The keys are allocator-chosen xmlNodePtr
addresses (not adversarial input), so HashDoS resistance is irrelevant and SipHash
is pure wasted work. Swapping in a dependency-free FxHash-style multiply-rotate
hasher on the pointer key cuts ~28-30% of wall on math/node-heavy documents
(latexml-oxide: 1510.03361 19.6s→14.1s, a tikz-cd paper 22.4s→15.7s, across both
conversion phases).

• Zero new dependencies (std::hash only).
• The map is never iterated, so losing per-run randomization is a no-op.
• Output-identical.
• Includes a determinism + no-collision regression test.

2. perf(attrs): get_properties reads values directly from the attr node

get_properties / get_properties_ns (and the get_attributes aliases) looped
over the attribute list but then re-resolved each value by name via
get_property → xmlGetProp, which re-scans the whole attribute list
(xmlStrEqual per entry) and allocates a fresh CString for the name on every
attribute — quadratic in the attribute count, and pure overhead since the loop
already holds the attribute node pointer.

This reads the value straight from that node with xmlNodeGetContent (the same
call get_content already uses), via a small attr_node_value helper. The
returned map is unchanged; libxml's own attribute-accessor tests pass.

Validation

cargo test --release — all test binaries pass (0 failed), including the new
node-cache determinism test.