Send AI spans as v2 spans

[sentrivana]

We have higher per-span size limits for the new v2 spans than we have in our transaction-based event payload.

Today, many AI spans (and with them their whole transaction) are dropped because they can exceed size limits. We try to alleviate this via truncation, but this a) looses information and b) is not perfect and misses certain things, so the situation is still not ideal.

To fix this, we want to send such AI spans in a combined envelope:

If a transaction should be sent and we detect it contains AI spans,
… then we want to take the AI spans out of the spans array of the transaction event…
… and instead send these as v2 spans in a separate envelope item in the same payload

This should ensure that AI spans are not dropped.

Note: This means that on e.g. self hosted instances that do not support v2 spans things may be lost, but this is an acceptable downside for the time being.

Description shamelessly stolen from getsentry/sentry-javascript#20170

[linear-code]

PY-2294

[sentrivana]

Some notes on a possible way to implement this. This can be quick and dirty since it'll go away in the next major where span first will be the default.

• Add a feature flag (see what it's called in JS to align on a name) to _experiments.
• If the flag is True:
  • In the place where we currently construct the transaction envelope, iterate through the spans it contains and if they're gen_ai, remove them from the envelope.
  • Take the removed spans and turn them into the V2 format
    • When translating to v2, we need to map all data to attributes. These should be in semantic conventions.
  • Put the new spans in a new envelope item and send as usual.