fix(serialization): surface collected deserialization errors instead of masking them

[spichen]

Problem

PydanticComponentDeserializationPlugin.deserialize rebuilds the collected validation errors into pydantic_core.InitErrorDetails carrying only (type, loc):

InitErrorDetails(type=e.type, loc=e.loc, input=())

ValidationError.from_exception_data then re-derives each builtin type's required ctx (value_error needs {"error": <exc>}, gt needs gt, …). With none supplied, a collected value_error — any component model_validator that raises ValueError — makes from_exception_data itself raise:

TypeError: 'error' required in context

…which masks the real cause. Seen in the field as a Swarm-used-as-a-Flow-node AgentNode port mismatch (ValueError("... expected a property titled 'evidence_status' ...")) surfacing only as the opaque TypeError: ValueError: 'error' required in context.

Fix

Rebuild each collected error with a PydanticCustomError(e.type, e.msg), which carries the message directly — rendered verbatim (no ctx, no template interpolation). This reconstructs an error of any type without the per-type ctx dance, and preserves both the original type and msg.

Notably simpler than special-casing value_error/ctx: no branch on type, no message munging, no double-prefixing.

Tests

tests/serialization/test_deserialization_error_surfacing.py:

• a collected value_error re-raises as a ValidationError (not TypeError) with its message verbatim (no added/doubled Value error, prefix);
• the original error type is preserved (a missing stays missing, not flattened to value_error);
• messages carrying literal {/} (JSON) survive verbatim (not treated as a format template);
• multiple errors all surface, error_count == 2.

Full tests/serialization/ suite: 303 passed (the 4 pre-existing test_llm_node.py default_llm_config fixture errors are unrelated to this change).