sample: maintain token count in penalty sampler context

[kylo5aby]

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[ggerganov]

Need to also reset token_count in llama_sampler_penalties_reset()

[kylo5aby]

Need to also reset token_count in llama_sampler_penalties_reset()

Resolved.

[ggerganov]

I think an assert here would be worth it as a sanity check:

Suggested change

	if (ctx->prev.size() >= (size_t) ctx->penalty_last_n) {
	if (--ctx->token_count[ctx->prev.front()] == 0) {
	if (ctx->prev.size() >= (size_t) ctx->penalty_last_n) {
	assert(ctx->token_count.at(ctx->prev.front()) > 0);
	if (--ctx->token_count[ctx->prev.front()] == 0) {