Optimize time-series source operator (elastic#127095)

This query against the TSDB track took 50 seconds and was reduced to 19 seconds with this changes.

```
TS tsdb 
| STATS sum(rate(kubernetes.container.memory.pagefaults)) by bucket(@timestamp, 5minute)
```

This change introduces several optimizations to improve the performance 
of the time-series source operator:

- Split the leaf queue into two: one for `_tsid` and another for `@timestamp`. 
This avoids repeatedly comparing large `_tsid` values while iterating over a single `_tsid`.

- Track the number of emitted documents per segment and use this data to 
  build forward and backward document maps, reducing the need for
  expensive sorts.

- Use ordinal blocks to avoid duplicating the same `_tsid` multiple times

Author: dnhatn

x-pack/plugin/esql/compute/src/main/java/org/elasticsearch/compute/data/DocVector.java

@@ -67,6 +67,12 @@ public DocVector(IntVector shards, IntVector segments, IntVector docs, Boolean s
         blockFactory().adjustBreaker(BASE_RAM_BYTES_USED);
     }
 
+    public DocVector(IntVector shards, IntVector segments, IntVector docs, int[] docMapForwards, int[] docMapBackwards) {
+        this(shards, segments, docs, null);
+        this.shardSegmentDocMapForwards = docMapForwards;
+        this.shardSegmentDocMapBackwards = docMapBackwards;
+    }
+
     public IntVector shards() {
         return shards;
     }
@@ -209,8 +215,12 @@ protected void swap(int i, int j) {
         }
     }
 
+    public static long sizeOfSegmentDocMap(int positionCount) {
+        return 2 * (((long) RamUsageEstimator.NUM_BYTES_ARRAY_HEADER) + ((long) Integer.BYTES) * positionCount);
+    }
+
     private long sizeOfSegmentDocMap() {
-        return 2 * (((long) RamUsageEstimator.NUM_BYTES_ARRAY_HEADER) + ((long) Integer.BYTES) * shards.getPositionCount());
+        return sizeOfSegmentDocMap(getPositionCount());
     }
 
     @Override