Upgrade protobuf from v4 to v5: 5x Upsert throughput (#393)

Upgrade the protobuf dependancy from v4 (4.25) to v5 (5.28). This
appears to have significantly faster protobuf encoding - I see a 4.5x -
5x
inprovement in Upsert throughput on a given EC2 machine (i3.xlarge)
for large batches (~300) of high dimensionality vectors (1536):

Before:

Performing Populate phase 1675770/138364198
╸━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1% 0:35:34 44:17:17
      Records/sec: 785.2

After:

Performing Populate phase 1531830/138364198
╸━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1% 0:07:07 10:35:48
      Records/sec: 3584.4

I haven't dug into the exact details, but the profile is quite
different - the python frames performing type checking are no longer
present, so I assume they have been optimised, perhaps pushed to
native code?

Before profile:


![protobuf_v4](https://github.com/user-attachments/assets/2c201489-c3f0-489b-8db1-e393c8953747)

After profile:


![protobuf_v5](https://github.com/user-attachments/assets/608ed8a0-9140-469f-88cc-d5546195050c)

## Type of Change

- [x] None of the above: Dependency upgrade.

## Test Plan

Describe specific steps for validating this change.

pyproject.toml

@@ -49,7 +49,7 @@ grpcio = [
 ]
 googleapis-common-protos = { version = ">=1.53.0", optional = true }
 lz4 = { version = ">=3.1.3", optional = true }
-protobuf = { version = "^4.25", optional = true }
+protobuf = { version = "^5.28", optional = true }
 protoc-gen-openapiv2 = {version = "^0.0.1", optional = true }
 pinecone-plugin-interface = "^0.0.7"
 pinecone-plugin-inference = "^1.1.0"