rename `max_tokens` throughout to `max_generation_tokens`

[leondz]

So distinct from max input tokens, max context window.

See discussions in #1112

[leondz]

#1112 (comment)

consider garak values:

• ctx_len - total context window size in tokens
• max_input_tokens - maximum input/prompt size for a generator
• max_output_tokens - maximum available/requested output size (less is fine, this is a cap not a demand)
• input_overhead_tokens - fixed costs on inputs (see OpenAI chat modality)
• max_prompt_tokens - max prompt length this turn, given overhead, system prompt/conv history, ctx_len, max_input_len, and "some output" (suggest min 150 tokens)

we assume tiktoken by default, and I guess all of these params are optional

[jmartin-tech]

I would expand on this issue to contextualize that removal of max_tokens is not the primary goal. Accounting for and enforcing token budgets needs to be more straight forward and consistent for all generators.

[leondz]

Oh, is it not? I see accounting and naming as parallel, independent efforts

[jmartin-tech]

The naming to me is intrinsically linked to how it is used, accounting from a feature vector is independent to some extent but core usage and name are coupled in my mind at this time.

[leondz]

OK. I am mindful of oai-specific maths affecting how we think of token counting in general.