fix: make return/throw behavior setters fully mutually exclusive

[rajanpanth]

Summary

Each of returns/returnsArg/throwsArg/returnsThis/callsFake/resolves/resolvesArg/rejects/resolvesThis/callThrough/callThroughWithNew only clears the specific fields its author happened to think of, rather than the full set behavior.js#invoke checks. So e.g. returnsArg followed by throwsArg, callsFake, returnsThis, or resolves leaves returnArgAt in place, which invoke() checks before any of those — silently keeping the earlier behavior instead of being overridden:

const stub = sinon.stub();
stub.returnsArg(0);
stub.throwsArg(1);
stub("myarg", new Error("boom")); // returns "myarg" -- does NOT throw

#d36e921 fixed this for the returns-after-returnsArg pair specifically, but the same bug shape remained for every other combination.

Changes

• src/sinon/default-behaviors.js: generalizes that fix into a shared resetBehavior(), called by every return/throw setter before it sets its own field(s). This is the same exhaustive-reset pattern callThrough already used — I just made every setter in this family use it consistently instead of each remembering to clear a handful of fields ad hoc.
  • Deliberately excludes the callsArg/yields callback-invocation fields from this reset: yielding a callback is a separate, additive side effect meant to combine with any of these return/throw behaviors regardless of call order (see the existing "returnsArg takes precedent over yielded return value" tests) — not another mutually-exclusive choice in this set. Left the callsArg/yields family of setters untouched.
  • callThrough/callThroughWithNew still additionally clear the callback config too, since "call the real method" is meant to replace all stub customization. This also fixes callThroughWithNew never having cleared returnArgAt/returnThis/fakeFn/resolve*/reject/returnValue, unlike callThrough (which already cleared everything).
• test/src/stub-test.js: 6 new "is superseded by X" tests alongside the existing ones, covering the previously-broken combinations.

Testing

• Full suite: 1503/1504 passing. The 1 failure (Sandbox > warns of potential leak when many fakes are created, a 2000ms timeout) is pre-existing and unrelated — confirmed via git stash, it fails identically without this change.
• Confirmed all 6 new tests fail on the prior code and pass with the fix.
• eslint --max-warnings 0 clean.