Order the ids in the order they were defined in codegen_llvm.h. This
makes it easier to spot missing ids.

We were missing cgroup_path_id and watchpoint_id. Without this, there
would be bugs for cgroupid() if it is used inside a loop or in an
expanded probe. Similarly, if watchpoint probes are expanded, there
would have been bugs.

There is no need to keep the strlen value explicitly on stack. We can
leave it up to LLVM whether or not to spill onto stack or keep in a
register.

Before, strncmp codegen assumed that caller was passing in two alloca'd
strings. This assumption was fine but overly narrow now that we are
implementing big strings. Big strings are not an array type (alloca),
but rather a proper pointer.

Relax assumption by using a more general pointer to i8.

Previously, all strings needed to be stored on the stack. Given that BPF
has a 512B stack limit, this put serious limits on strings. In practice,
it was limited to about 200B.

This commit raises the limit much higher to around 1024. The new limit
we hit is from the LLVM memset() builtin. It only supports up to 1K. To
go above this, we need to create our own memset() routine in BPF. This
is quite easy to do and will be attempted in a later commit.

The way big strings work is by assigning each str() call in the script a
unique ID. Then we create a percpu array map such that for each CPU,
there is a percpu entry that can accomodate each str() callsite. This
makes it so scripts can keep as many strings "on the stack" as they'd
like, with the tradeoff being the amount of memory we need to
preallocate.

As long as the kernel never allows BPF programs to nest (eg execute prog
A while prog A is already on the call stack), this approach is robust to
self-corruption.

Note we do not yet remove or change default value of
BPFTRACE_MAX_STRLEN. That is b/c a few other helpers (buf(), path(),
maybe more) also read the value and allocate space on stack accordingly.
We need to decouple or convert those helpers to scratch maps as well.

This closes bpftrace#305.