…und turns

The width checks used the 5th to 95th percentiles of the points, which
read an evenly spread object about 10% short: a 9 cm block measured
8.1 cm and the turned grasp brought 8.5 cm jaws down on it. Widths are
now measured between the 1st and 99th percentiles, and a grasp is
offered only when the object is narrower than the jaw opening by
jaw_clearance (0.01 m, config). The same holds for each half.

The turned grasp is also skipped when the footprint's length and width
differ by less than 1 cm: on a round or square object it holds the same
way as the first grasp and would only use up a retry.

Move HeuristicGraspModule's grasp math into heuristic_grasps(cloud,
config), with the grasp rules in HeuristicGraspConfig. The module's
propose_grasps calls it and returns the same grasps as before, so other
modules can compute heuristic grasps without a second grasp provider.

…M1 pick

Each grasp try now plans to the provider's best max_grasp_attempts new
candidates, then to heuristic grasps computed from the same points. The
re-scan after a miss proposes both again, and a candidate from either
source that repeats a grasp already tried is skipped. A GraspGenX pick
then still has grasps when GraspGenX returns none, raises, or its
grasps cannot be reached or miss within the retry budget.

- A provider error is logged and treated as no provider grasps; the try
  records it as grasp_provider_error.
- Heuristic grasps within retry_skip_distance and retry_skip_angle of a
  provider grasp being tried are dropped; a provider that is
  HeuristicGraspModule gets no second copy.
- Every pick result carries grasp_sources; each try records its source
  and candidates_by_source.
- PickAndPlaceModuleConfig.heuristic_fallback (default True) turns it
  off, giving the provider's grasps only and raising its errors as
  before. heuristic_grasp sets the gripper fit.
- _repeats uses the shared same_grasp check.

…e M1 pick, report rejections

GraspGenX's best-scoring xArm grasps on SceneReplica objects are mostly side
grasps at table height, many approaching from below or from the far side.
The planner's Pink IK rejects those in about 0.25-0.3 s each (10 restarts of
200 iterations), so with only 5 tried each try nearly always fell through to
the heuristic grasp. max_grasp_attempts now defaults to 25.

Each record in tries now carries rejected_by_source (how many grasps from each
source the planner found no path to in that try), and the pick result carries
the total over the whole pick, including a final try that reached no grasp. A
move to a grasp that stops part-way is now recorded as a try too, so its
source and rank are reported.