Improve GridQubit dictionary lookup performance 3x (#3375)

According to the following profiling code, this change decreases the dictionary lookup time from 2100ns to 600ns on my machine.

```
import time
import cirq

d = {q: 5 for q in cirq.GridQubit.rect(3, 3)}
q = cirq.GridQubit(0, 0)
total = 0
t0 = time.monotonic()
n = 100000
for _ in range(n):
    total += d[q]
t1 = time.monotonic()
print(total)
print((t1 - t0) / n * 10**9, "nanoseconds per hash")
```

Author: Strilanc

cirq/devices/grid_qubit.py

@@ -275,6 +275,26 @@ class GridQubit(_BaseGridQid):
         cirq.GridQubit(5, 4)
     """
 
+    def __init__(self, row: int, col: int):
+        super().__init__(row, col)
+        self._hash = super().__hash__()
+
+    def __hash__(self):
+        # Explicitly cached for performance (vs delegating to Qid).
+        return self._hash
+
+    def __eq__(self, other):
+        # Explicitly implemented for performance (vs delegating to Qid).
+        if isinstance(other, GridQubit):
+            return self.row == other.row and self.col == other.col
+        return NotImplemented
+
+    def __ne__(self, other):
+        # Explicitly implemented for performance (vs delegating to Qid).
+        if isinstance(other, GridQubit):
+            return self.row != other.row or self.col != other.col
+        return NotImplemented
+
     @property
     def dimension(self) -> int:
         return 2