fix(iter): map-left/map-right iterate the named operand

[singaraiona]

Problem

map-left/map-right were defined by which argument is held fixed, the inverse of the conventional model. This caused surprising results whenever the scalar sat in the "wrong" slot or the function was a non-broadcasting lambda:

(map-right (fn [x y] (println "x=% y=%" x y)) 10 [1 2 3])   ;; gave one call, expected three
(map-right concat 10 [1 2 3])                                ;; gave [10 1 2 3], expected ([10 1] [10 2] [10 3])

A prior attempt removed an "auto-detect" heuristic, which only masked the issue because the test suite exercised the cases where atomic builtins broadcast either way.

Fix

Define the ops by which operand is iterated:

• map-left — iterate the left, hold the right whole → fn(left_i, right)
• map-right — iterate the right, hold the left whole → fn(left, right_i)

The iterated side is fixed by the operator (no auto-detect). When that side is an atom there's nothing to iterate, so fn is applied once to (left, right) — and for an atomic builtin that single call is its broadcast, so map-left/map-right agree with plain application there. Every call routes through call_fn2, which already broadcasts atomic builtins over a held vector and hands lambdas their arguments whole.

(map-right (fn [x y] (list x y)) 10 [1 2 3])  → ([10 1] [10 2] [10 3])
(map-right concat 10 [1 2 3])                  → ([10 1] [10 2] [10 3])
(map-left  + 10 [1 2 3])                       → [11 12 13]
(map-right - [10 20 30] 5)                     → [5 15 25]

Migration

This swaps the prior meaning of the two ops. The window-join interval idiom built its N per-row [lo hi] intervals with the old map-left; that role is now map-right, so the join tests and docs are migrated map-left → map-right (verified to produce identical window-join results). Reference docs for the two ops are updated, and non-broadcasting lambda regression tests were added (a concat / list-building fn exposes the iteration structure that arithmetic builtins' broadcast otherwise hides).

Tests

Full suite green: 3443/3445 passed (2 skipped, 0 failed).