fix(query): reject STR equality keys in asof-join and window-join

[belowzeroff]

Problem

asof-join, window-join, and window-join1 accept string (STR) equality-key columns and then return silently wrong results — no error, corrupt data.

Root cause: all three read equality-key cells through read_col_i64, which has no RAY_STR case, so a STR column falls into the byte-wide default (((uint8_t*)data)[row]).

• asof-join (asof_eq_lread) mismatches outright — a key that should match nulls out:

  ;; identical data, only the key TYPE differs
  (asof-join [Sym Time] tr qs)  with SYM Sym -> Bid [99 99]   correct
  (asof-join [Sym Time] tr qs)  with STR Sym -> Bid [99 0Nl]  WRONG
  

• window-join / window-join1 read one byte per row when they sort and probe the right side. Small key sets coincide by luck, but once there are more distinct keys than fit in a byte the collisions cross-contaminate groups:

  ;; 300 distinct keys, each with its own quote; per-key minBid must equal the key
  (sum (- minBid (til 300)))  with SYM keys ->      0   correct
  (sum (- minBid (til 300)))  with STR keys -> -41605   WRONG
  

Fix

The base inner-join / left-join / anti-join use a separate STR-aware kernel (ray_join / ray_antijoin) and stay correct even with 300 distinct STR keys (verified), so this is scoped to the two window/asof kernels only.

Decline a STR equality key on either side with nyi instead of returning corrupt data (parted columns unwrapped to their base type first, matching the base join). One guard in ray_asof_join_core covers both asof call paths (plain + parted/day); one in window_join_impl covers both window-join and window-join1.

nyi (rather than adding STR support to these kernels) is the conservative bug fix: it turns silent corruption into an explicit, catchable error. Real STR eq-key support in the window/asof kernels can be a follow-up.

Tests

test/rfl/join/str_key_nyi.rfl:

• STR eq-key -> nyi for asof-join (both operand orders), window-join, window-join1;
• SYM eq-keys keep working and stay order-independent on the right side;
• base inner-join / anti-join with STR keys must not be over-rejected (regression guard).

make test: 3673 of 3674 passed (1 skipped, 0 failed).

[singaraiona]

Blocking: the new STR guard does not cover the supported 4-argument compatibility form of window-join.

The guard is inside the n >= 5 Rayforce-convention branch of window_join_impl, so (window-join left right [eq-keys] time) bypasses it and falls through to ray_asof_join / exec_window_join. I reproduced this on PR head da97945 with the same STR fixture:

(at (window-join trades quotes [Sym] 'Time) 'Bid)

It still returns [99 0Nl] rather than nyi, so silent corruption remains. The existing lang/eval/window_join test confirms this call shape is supported.

Please add the validation to the legacy fallthrough as well (or, preferably, reject STR after equality columns are resolved in the shared as-of executor), and add regressions for the 4-argument window-join and window-join1 paths. The new test currently exercises only the 5-argument window form, which is why the full suite still passes.

For context, the previous related fixes covered adaptive-width/cross-domain SYM; they did not add STR descriptor handling to read_col_i64.