Introduce annotations for large attributes

[lmolkova]

From discussion on #3521 (comment): it'd be useful to mark attributes that are large so that instrumentations, OTel SDK and consumers can process them differently.

The lightweight way of doing this is through annotations in semconv, they don't go over OTLP, but instrumentation / pipeline code can be written / generated in a way that handles these attributes differently (e.g. uploads to a different type of storage).

Proposed annotation name: value_size

Alternatives: value_size_hint, expected_value_size, value_size etc

This is how it looks like along with other existing or possible future annotations

annotations:
  value_size: large # enum: large, short; none means unknown / medium   
  cardinality: high # enum: low, medium, and high - useful to validate metric and identifying attrs definitions
  sensitivity:  PII # secrets, linked PII, etc - just a string for now. when present, it's sensitive in one of the ways
  # existing: 
  code_generation:
    exclude: true

[kamphaus]

Here are prior issues & pull requests that aim to deal with large attribute values:

• Seeking input on proposal for "reference"-type attribute values. #1428
• Add the concept of Blob Reference Properties ("reference attributes" and "reference fields") to the spec #1521
• Add gen_ai _ref attributes for referencing external uploaded content #2750

[Dipanshusinghh]

I like this direction. Looking at the earlier discussions in #1521 and #2750, it seems that handling large values through reference attributes introduced a lot of additional design questions around storage and retrieval.

Using an annotation feels much lighter while still giving instrumentations and backends enough information to treat certain attributes differently when needed.

One thing I'm wondering about: is there a reason to distinguish between short and large, instead of only marking attributes that are known to be large and leaving everything else unspecified?