[enhancement] reseed random number when loading from cache and --seed provided

[d-takemori]

Current Behavior

When loading from a saved --prompt-cache file, llama.cpp appears to initialize response generation with the same state that was saved with the cache-file. This yields the same output every run.

Initial run

% ./main --model ../../Models/LLaMA/7B/ggml-model-q5_1.bin --prompt-cache test-cachefile -n 36 --color -p 'What happens with a test prompt?'
main: build = 537 (6456a4e)
main: seed = 1684011989
[snip]
main: attempting to load saved session from 'test-cachefile'
main: session file does not exist, will create
[snip]
What happens with a test prompt?
A test prompt is an assignment that helps you find out whether something works.
This can be anything from finding out if your code does what it should do or if the data

next run

% ./main --model ../../Models/LLaMA/7B/ggml-model-q5_1.bin --prompt-cache test-cachefile -n 36 --color -p 'What happens with a test prompt?'
main: build = 537 (6456a4e)
main: seed = 1684012061
[snip]
main: attempting to load saved session from 'test-cachefile'
main: loaded a session with prompt size of 8 tokens
main: session file has exact match for prompt!
[snip]
What happens with a test prompt?
A test prompt is an assignment that helps you find out whether something works.
This can be anything from finding out if your code does what it should do or if the data

another run, but with explicit --seed

% ./main --model ../../Models/LLaMA/7B/ggml-model-q5_1.bin --prompt-cache test-cachefile -n 36 --color -p 'What happens with a test prompt?' --seed 1234
main: build = 537 (6456a4e)
main: seed = 1234
[snip]
main: attempting to load saved session from 'test-cachefile'
main: loaded a session with prompt size of 8 tokens
main: session file has exact match for prompt!
[snip]
What happens with a test prompt?
A test prompt is an assignment that helps you find out whether something works.
This can be anything from finding out if your code does what it should do or if the data

Desired Behavior

If loading from a cache-file AND "--seed" # is given, llama.cpp should initialize the random number generator before response generation, so that different responses are generated.

If loading from a cache-file and "--seed #" not given, preserve current behavior.

[aleksusklim]

I posted a comment about other strange behavior of prompt caching:
#1550 (comment)

UPD: reposted it as a separate issue: #1585