feat: add BB_POLY_STATS for prover polynomial memory analysis

[johnathan79717]

Summary

• Adds optional instrumentation (gated behind BB_POLY_STATS env var) that reports per-polynomial value-size distribution after ProverInstance construction
• Helps assess feasibility of variable-length encoding for memory reduction (re: Investigate zero-byte content in prover polynomials for memory compression barretenberg#1623)

Usage

BB_POLY_STATS=1 ./build/bin/chonk_bench --benchmark_filter=Full/2

Outputs a table like:

=== Polynomial Memory Analysis ===
Polynomial                           |  AllocSize |      Zeros |       <=4B |       <=8B |      <=16B |      <=32B |    Mem(MB) |  Compr(MB)
sigma_1                              |     336637 |          0 |     336629 |          0 |          0 |          8 |      10.27 |       1.28
w_l                                  |     336637 |      45462 |     165039 |      13556 |      76637 |      35943 |      10.27 |       3.00
...
TOTAL                                |    8125052 |    3192926 |    4004383 |     148196 |     318296 |     461251 |     247.96 |      35.34

Total actual memory:     247.96 MB
Total compressed memory: 35.34 MB
Potential savings:       85.7%

Initial findings (2^19 mock app circuit)

• 85.7% theoretical savings with ideal variable-length encoding
• Sigma/ID polys (8 polys, ~82 MB): 100% values fit in 4 bytes — uint32_t gives 8x compression
• Wire polys (w_l–w_4, ~41 MB): ~50% fit in 4B, ~20% need full 32B
• Selector polys (~96 MB): ~80% zeros, rest mostly small — already partially optimized via mega-selectors
• z_perm, lookup_inverses (~11 MB): 100% zeros at construction time (filled later)

Test plan

• Builds with cmake --build build --target chonk_bench
• BB_POLY_STATS=1 ./build/bin/chonk_bench produces correct per-polynomial report
• No output when BB_POLY_STATS is unset (zero overhead in normal operation)

[ledwards2225]

Very cool!