Fix potential pointer math truncation and improve readability

[davidfiala]

Fix potential pointer math truncation bug. Improve readability of datatypes when binary length or handling is at play.

• Fix potential pointer error due to integer truncation in read_checkpoint. Although unlikely due to compiler padding, *data + sizeof(Config)/sizeof(float); could can been fractional and thus truncated, leading to an incorrect pointer. Admittedly rare given the pointer padding, but given how people depend on llama2.c as a canonical learning resource, it should be fixed.
• Simplify reading of weights by opening the file only once. Less LOC, less opens by using just one fd now.
• Simplify code related to "data = start of file + sizeof(config)" in that same function
• Add packing pragmas to ensure that the compiler understands we're reading a binary file in mmap'd structs to avoid implicit padding anywhere in the structs. Although offsets of 4x bytes often works out of the box, I'd hate for people to copy/paste this code and in some case get unexpected results/corrupt pointers.
• Simplify/clarify datatypes for a few places where we're dealing with binary data and binary math such as Config and rng_seed: use uint64_t and [u]int32_t to make it absolutely clear what the binary format is and what length's we're dealing with rather than depending on compiler convention of int and unsigned long long
• improvement: parse CLI flag seed as a uint64_t rather than int, unlocking the full range of possible seeds rather than just those up to 2^31-1