feat(run): propagate the calling agent as an x-agent header (ENG-3488)

[elsheikhams99]

Runnable resources already carry the caller identity as x-user-id and the
run's session as x-session-id; add the calling agent's name so downstream
services can attribute a call to the agent that made it.

• agent_name -> x-agent, header-only for every runnable like session_id
  (no run params type declares it as a body field), so the base
  _RUN_CONTROL_KEYS excludes it and Model/Tool also strip it from the v1/URL
  builder paths. Declared on ModelRunParams, not BaseRunParams, for the same
  reason as session_id: AgentRunParams must stay free of look-alike keys.
• Collapse the header list into one source of truth,
  RunnableResourceMixin._RUN_HEADER_KEYS. _headers_for_run now iterates it and
  Model._SDK_ONLY_PARAMS derives its header half from it, instead of both
  re-listing the same kwargs.

The refactor is motivated by 9cba90d: the merge of 'test' into development
(dc8c4c2) resolved the _SDK_ONLY_PARAMS hunk in favour of the older frozenset,
dropping session_id while keeping its tests and comments -- so the value rode
in the request body until someone noticed. With the keys derived rather than
duplicated, the payload filters can no longer drift from the wire.

Adds unit coverage for x-agent emission and payload exclusion (v2 + v1
builders), all three headers together, and
TestRunHeaderKeysAreSingleSourceOfTruth -- which asserts every _RUN_HEADER_KEYS
kwarg is stripped from both payload paths, and would have caught the regression
above.

[aix-ahmet]

Reviewed together with aixplain/aixplain-agents#306. The _RUN_HEADER_KEYS single-source-of-truth refactor is the right call — it makes the 9cba90d class of merge regression structurally impossible, and TestRunHeaderKeysAreSingleSourceOfTruth proves it. Coverage is thorough. One change requested, one nit.

1. Non-ASCII agent_name surfaces as a deep UnicodeEncodeError at send time — validate and raise a clear error instead

Unlike identifier and session_id (platform-generated, ASCII-safe ids), agent_name is user-authored free text. The header builder passes it straight through:

aiXplain/aixplain/v2/resource.py

Lines 1304 to 1308 in db9440c

	for key, header in self._RUN_HEADER_KEYS:
	value = kwargs.get(key)
	if value is not None:
	headers[header] = str(value)
	return headers or None

Verified: requests/urllib3 accepts the header at prepare time but crashes at send with UnicodeEncodeError: 'latin-1' codec can't encode character 'ı'... for a name like "Ürün Asistanı" (any CJK/Arabic name fails too). That error gives the caller no hint that the agent's name is the problem.

Decision: fail fast with a descriptive error — do not sanitize or alter the name. In _headers_for_run (or a small validator it calls), reject a non-ASCII header value with a ValueError naming the offending kwarg and value, e.g. ValueError("agent_name 'Ürün Asistanı' cannot be sent as the x-agent header: header values must be ASCII"). Please add a test with a non-ASCII name asserting the error.

2. Nit: docstring team attribution

aiXplain/aixplain/v2/model.py

Lines 502 to 505 in db9440c

	agent_name: Name of the agent making this call, emitted as the ``x-agent``
	header so downstream services can attribute it to the calling agent
	(for a team run, the specific sub-agent — not the team). Header-only,
	on the same terms as ``session_id``.

"for a team run, the specific sub-agent — not the team" is true for delegated calls, but the orchestrator's own direct tool calls carry the root/team name. Worth a small wording tweak so consumers of the header aren't surprised.