fix: admit zero-bytes-saved samples so skip fires for filter==projection

Q18 (and several other TPC-DS regressions) had a post-scan
customer_demographics filter — `Optional(DynamicFilter)` on the single
projected column `cd_demo_sk` — that burned 90 ms of CPU per scan but
could never be skipped. The filter was correctly placed at PostScan
(projection ⊆ filter columns ⇒ byte_ratio = 1.0 > threshold) but the
mid-stream skip path never fired.

Root cause: `SelectivityStats::update` only incremented `sample_count`
when `batch_bytes > 0`. When the projection is a subset of the filter
columns, `other_bytes_per_row = 0` and therefore `batch_bytes = 0` on
every call, so the Welford counter stayed at zero, the CI upper bound
stayed `None`, and the skip check short-circuited. Meanwhile the filter
kept running per batch.

Admit samples with `batch_bytes = 0`. The recorded effectiveness for
those samples is legitimately zero (no late-materialization payoff), so
the CI upper bound converges on zero after a few batches and the skip
flag flips for optional filters — exactly what we want: CPU spent, no
byte savings, optional ⇒ drop.

Local TPC-DS sf1 (M-series, pushdown=on) vs main pushdown=off:

| Query | Main(off) | Before | After |
|-------|-----------|--------|-------|
| Q18   |   99      | 182    |  118  |
| Q67   |  312      | 503    |  346  |
| Q26   |   80      | 151    |   94  |
| Q85   |  149      | 246    |  157  |
| Q91   |   64      | 108    |   58  |
| Q53   |  103      | 144    |   99  |
| Q63   |  103      | 148    |   99  |
| Q13   |  399      | 558    |  376  |
| Q72   |  619      | 489    |  277  |
| Q24   |  379      |  70    |   70  |
| Q64   | 28213     | --     |  519  |

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: adriangb

datafusion/datasource-parquet/src/selectivity.rs

@@ -149,10 +149,20 @@ impl SelectivityStats {
         self.eval_nanos += eval_nanos;
         self.bytes_seen += batch_bytes;
 
-        // Feed Welford's algorithm with per-batch effectiveness
-        if total > 0 && eval_nanos > 0 && batch_bytes > 0 {
+        // Feed Welford's algorithm with per-batch effectiveness. We admit
+        // samples with `batch_bytes == 0` — that legitimately represents a
+        // filter whose projection is a subset of its referenced columns, so
+        // late materialization has nothing to save even when the filter
+        // does prune rows. Recording `batch_eff = 0` for such batches lets
+        // the mid-stream skip path detect "CPU spent, no late-
+        // materialization payoff" and drop the filter if it is optional.
+        if total > 0 && eval_nanos > 0 {
             let rows_pruned = total - matched;
-            let bytes_per_row = batch_bytes as f64 / total as f64;
+            let bytes_per_row = if total > 0 {
+                batch_bytes as f64 / total as f64
+            } else {
+                0.0
+            };
             let batch_eff =
                 (rows_pruned as f64 * bytes_per_row) * 1e9 / eval_nanos as f64;