feat: add theta_star, an any-angle search

[tachsin]

Draft for #829 — something concrete to react to, not a finished proposal. If the answer to question 1 over there is "no", close this and nothing is lost.

What it does

astar with one changed relaxation step. On reaching a node, the search first asks whether the node it came from can see the successor directly; if so the successor is attached to that node by a straight line instead of stepping through the current one. The waypoints stay graph nodes, the segments between them do not.

pub fn theta_star<N, C, FN, IN, FH, FL, FS>(
    start: &N,
    successors: FN,
    heuristic: FH,
    sight: FL,     // FnMut(&N, &N) -> Option<C>
    success: FS,
) -> Option<(Vec<N>, C)>

sight gives the cost of a straight line between two nodes that need not be neighbours, or None when it is blocked. That is the only thing Theta* needs that astar does not, so line-of-sight and the straight-line metric both stay with the caller and nothing here gains a dependency on coordinates or floating point.

150 lines of algorithm, 207 of tests.

The two caveats from the issue, as implemented

Both are documented on the function under their own heading rather than buried:

• Not optimal. Never longer than the astar path, usually close to the true any-angle optimum, but no guarantee — that needs a visibility graph. First inexact search in the crate.
• Consecutive path nodes need not be neighbours. path.windows(2) are straight lines, not edges of successors.

There is also a stated requirement on the caller: sight has to agree with successors in the triangle-inequality sense, or the shortcut can make a path worse and the reported cost stops matching it.

Testing

Three properties, on 8-connected grids with randomly blocked cells:

• With sight always returning None, the result equals astar exactly — path and cost, over 40 random maps. The shortcut is the only difference, so refusing every shortcut must reproduce astar.
• On open ground the path is [start, goal] at exactly the straight-line cost.
• On 60 random maps: never longer than astar, every segment is a clear line, the segment lengths sum to the cost reported, and no waypoint repeats. The test also asserts it is shorter than astar on most maps, so a version that silently stopped taking shortcuts would fail rather than pass quietly.

Plus a walled-off goal returning None.

296 tests pass, clippy and rustfmt clean.

One thing I would flag for review

SmallestCostHolder is shared with astar rather than duplicated — it becomes pub(super) with pub(super) fields. No public API change, but it does mean astar.rs is touched by this. Happy to duplicate the 30 lines instead if you would rather the modules stayed independent.

Still open from #829: whether sight should be one closure or two, whether a line-of-sight helper on Grid would be welcome, and whether Lazy Theta* is worth having alongside. None of those are settled here.