perf(mpt): specialize common witness RLP forms

[Qumeric]

Stacked on #658.

Specialize the dominant MPT witness encodings during decoding: empty branch-child references and fixed 32-byte unresolved digest nodes. The shared RLP-item helper also removes repeated header bookkeeping across leaf, extension, and branch decoding, while uncommon encodings continue through the generic alloy_rlp decoder.

A digest-root round-trip test covers the fixed unresolved-node representation.

Benchmark results

Block 24001988, RVR execution, comparing the #658 baseline 29744549529 with the RLP-only run 29811228788:

metric	#658	this PR	delta
execute_metered_insns	651,377,749	644,085,446	−7,292,303 (−1.12%)
metered_rows_unpadded	934,824,479	927,550,243	−7,274,236 (−0.78%)
metered_main_cells_unpadded	47,673,439,230	47,415,274,554	−258,164,676 (−0.54%)
metered_interaction_cells_unpadded	14,613,892,025	14,470,176,137	−143,715,888 (−0.98%)

Block 24003902

RVR execution: #658, this PR.

metric	#658	this PR	delta
execute_metered_insns	513,794,871	512,092,682	−1,702,189 (−0.33%)
metered_rows_unpadded	694,528,733	692,838,316	−1,690,417 (−0.24%)
metered_main_cells_unpadded	30,209,487,536	30,149,424,975	−60,062,561 (−0.20%)
metered_interaction_cells_unpadded	12,028,309,965	11,994,791,134	−33,518,831 (−0.28%)
metered_memory_unpadded_bytes	575,999,821,960	575,266,433,578	−733,388,382 (−0.13%)
memory-triggered segment cuts	51	51	0

[Qumeric]

Collapsed into #684 together with the other two RLP PRs. They were one idea split across three stacked PRs; #684 carries the union plus an rlp module and differential tests against alloy_rlp.