OpenVINO Backend: Upgrade graph cache to use buffer pointers instead of names

[mostafafaheem]

Overview

This PR replaces using names for graph_key struct with buffer pointers+offsets. This was done due to the previous approach sometimes causing false negatives, where the first decode graph would be deemed a cache miss, even though it is the same as the prefill graph.

Additional information

Example:

diff --git a/../prefill_graph.txt b/../decode_graph.txt
index 18ed618be..b6822c19c 100644
--- a/../prefill_graph.txt
+++ b/../decode_graph.txt
@@ -15,14 +15,14 @@ Graph input: 6:1:node_5#5502
 Graph input: 7:0:node_6#5594
 Graph input: 8:0:leaf_7#905
 Graph input: 9:0:node_7#5617
-Graph input: 9:1: (view)#5686
+Graph input: 9:1:leaf_7 (view)#5686^M
 Graph input: 10:0:leaf_8
 Graph input: 10:1:node_5#5502
 Graph input: 11:0:node_10#5640
 Graph input: 11:1:leaf_9
 Graph input: 12:0:leaf_10#928
 Graph input: 13:0:node_11#5663
-Graph input: 13:1: (view)#5709
+Graph input: 13:1:leaf_10 (view)#5709^M
 Graph input: 14:0:leaf_17
 Graph input: 14:1:node_5#5502
 Graph input: 15:0:node_14#5525

The views in the first decode graph accessed the leaf_7 and leaf_10 tensors after they were named, whereas they were still unnamed when the prefill graph accessed them. This is purely a name difference, the underlying graph is the exact same.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES, AI was used to explain relevant code relating to constructing graph keys, and help implementing an alternative using buffer pointers.