Add sample-based program generation to the fuzzer

[matteB10]

Extends the Trieste fuzzer so it can build random ASTs out of fragments taken from real sample programs. Sample subtrees are fast-forwarded to the start pass and updated for every subsequent pass when not running with --sequence mode. Experiments show that using samples reduces the number of retries to generate hash unique trees and increases code coverage.

New driver options:

• --samples takes a one or multiple paths to sample files.
• --sampling-level: 0 copies a full sample subtree; 1..n copies that many levels of a sample node and lets the generator fill in the rest.
• `--sampling-frequency, 0–100: how often a sample is used vs. random generation. 0 = never, 100 = always. Defaults to 50%. Stops generating if the tree reaches the ceiling depth.

Example usage for a driver-based reader:
infix_trieste test --samples examples/*.infix --sampling-level 1 --sampling-frequency 50

It is also added for the non-driver based json_fuzzer and yaml_fuzzer, but where the --samples option accepts a single path (possibly a directory).