Fix json_tool_call filtering: streaming support, pop->get, behavioral parity

[jquinter]

Summary

Addresses the 3 findings from the review of PR BerriAI#18384 (BerriAI#18384):

• Streaming support (requested by @krrishdholakia): AWSEventStreamDecoder now accepts json_mode and suppresses json_tool_call chunks in streaming, converting its arguments to text content. Updated all callers (make_call, make_sync_call in both invoke_handler.py and converse_handler.py) to pass json_mode to the decoder.
• optional_params.pop("json_mode") -> .get() (flagged by Greptile): Avoids mutating the caller's dict, which could break downstream logging/retries/metrics that access optional_params after response transformation.
• Preserve original tool_calls behavior: _filter_json_mode_tools() returns None only when json_tool_call was the sole tool (converted to content). Returns the list (including []) otherwise, preserving the original behavior of always setting tool_calls.

Also extracts the filtering logic into _filter_json_mode_tools() and moves import json to module level per project conventions.

Test plan

• test_transform_response_with_both_json_tool_call_and_real_tool — non-streaming mixed tools
• test_transform_response_does_not_mutate_optional_params — verifies .get() fix
• test_streaming_decoder_filters_json_tool_call — streaming with mixed json_tool_call + real tool
• test_streaming_decoder_without_json_mode_passes_all_tools — streaming passthrough when json_mode=False
• All 61 existing tests in test_converse_transformation.py pass