Prune placeholders only when a null was inlined

[gmpassos]

Follow-up to #152, which introduced a cost this removes.

The regression

1.15.1 rewrites field == ? bound to null into field IS NULL. That leaves the parameter resolved but unreferenced, and PostgreSQL rejects a statement carrying variables it does not use — so #152 added a prune that materialized the encoded output and scanned it for every placeholder:

var output = context.outputString;          // StringBuffer.toString(), every encode
parametersPlaceholders.removeWhere(
  (key, _) => !_isPlaceholderInOutput(output, parameterPlaceholder(key)));

Every query paid it, including the overwhelmingly common one that compares nothing against null — O(placeholders × output length), plus a '@$key' allocation per placeholder, plus the full toString().

Measured in isolation on a 3-placeholder condition of the shape entity queries produce:

	per encoded condition
1.15.1	0.376 us
this PR	0.074 us
removed	0.302 us

For scale, the 1.15.0 dispatch work brought a whole logged route call to ~0.90us. This was adding roughly a third of that back, on every query.

The fix

Only a placeholder actually rewritten to IS NULL/IS NOT NULL can become unreferenced. EncodingContext now records those keys as they are written, so:

• nothing compared against null → the prune returns immediately: no toString(), no scan, no allocation;
• something was → only the recorded keys are checked, and the output is materialized lazily, since a recorded key may still be referenced by another operator (field > ? bound to null) and then there is nothing to look for.

The set is left null while empty — allocating one per encoded condition would give back part of the saving.

Verification

No behaviour change is intended: same statements, same bound parameters. Checked against the suites that caught the original bug, where an unreferenced parameter is a hard error (Contains superfluous variables):

• PostgreSQL 65/65, MySQL 65/65 (Docker)
• Full VM suite 915 passing, ensure_build included
• dart analyze --fatal-infos --fatal-warnings and dart format --set-exit-if-changed clean

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 68.24%. Comparing base (c05f56e) to head (339ee9d).

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #153      +/-   ##
==========================================
- Coverage   68.24%   68.24%   -0.01%     
==========================================
  Files          66       66              
  Lines       22147    22160      +13     
==========================================
+ Hits        15114    15122       +8     
- Misses       7033     7038       +5     

Flag	Coverage Δ
unittests	68.24% <100.00%> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.