Add Trajectory.called_before?/4 for relative tool-call ordering assertions

[brainlid]

Summary

Adds LangChain.Trajectory.called_before?/4 and the companion
assert_called_before/4 / refute_called_before/4 macros for asserting
relative tool-call ordering — "tool A was called before tool B" —
independent of how many other calls happen in between.

This fills the gap between the existing modes:

• :superset enforces containment but is order-independent.
• :strict requires the whole sequence with exact count.

Relative ordering ("search before answer", "validate before write") is the
common middle ground for agent trajectory evals, and previously required
hand-rolling index math over tool_calls in every eval.

API

Trajectory.called_before?(trajectory, "search", "answer")
#=> true

assert_called_before trajectory, "search", "answer"
refute_called_before trajectory, "write_file", "read_file"

Accepts a Trajectory, an LLMChain, or a bare list of tool-call maps.

Semantics decisions

These resolve the open questions from the issue:

1. Which occurrences — uses the recommended "any A precedes any B"
   reading: min(index of A) < max(index of B). This tolerates interleaving
   (A…B…A…B) while still catching a pure B-before-A ordering.
2. Missing tools — by default a missing tool returns false (so
   refute_called_before passes vacuously). Passing require_both: true
   raises an ArgumentError naming the absent tool, so a missing tool is
   detected rather than silently collapsing to false. This mirrors the
   module's existing convention of raising ArgumentError on misuse in
   matches?/3, and applies symmetrically to both assert_ and refute_.
3. Same turn — ordered by flat-list position in tool_calls, consistent
   with the existing helpers.
4. Args matching — out of scope for v1; name-only, as suggested.

Tests

25 new tests covering the predicate (ordering, missing tools, interleaving,
repeated calls, LLMChain/list inputs, :require_both raising) and both
assertion macros. Full mix precommit is green (1799 tests + 31 doctests, 0
failures).

Closes #564