Features/streaming penalty

[p-ferreira]

• refactors synapse stream result
• add streaming data to logs and forward flow
• refactors protocol to return string chunks in streaming rather than string arrays
• adapt streaming processor to calculate tokens by chunk
• drops legacy prompting synapse from repo
• adds streaming reward function
• adds a global penalty definition to tasks
• adds unit tests

[dbobrenko]

LGTM

[dbobrenko]

Minor suggestion is to measure with perf_counter which is more precise and not affected by system clock

Suggested change

	start_time = time.time()
	# Calculate the accumulated penalty for the current chunk
	accumulated_penalty = sum(
	penalty_per_exceeding_chunk if tokens_per_chunk > self.max_tokens_per_chunk else 0
	for tokens_per_chunk in response_tokens_per_chunks
	)
	# Record the timing for this computation
	timings.append(time.time() - start_time)
	start_time = time.perf_counter()
	# Calculate the accumulated penalty for the current chunk
	accumulated_penalty = sum(
	penalty_per_exceeding_chunk if tokens_per_chunk > self.max_tokens_per_chunk else 0
	for tokens_per_chunk in response_tokens_per_chunks
	)
	# Record the timing for this computation
	timings.append(time.perf_counter() - start_time)