perf: reduce fetcher and evaluator calls in MeshTxBuilder

[lisicky]

What

Cuts the provider calls MeshTxBuilder.complete() makes, and fixes two builds that could not complete at all.

Contents

mesh-transaction (MeshTxBuilder):

• evaluateRedeemers skips evaluateTx and the serialization before it when nothing is witnessed by a Plutus script. Before, it ran once per coin-selection cost estimate on every transaction, including plain ADA sends.
• Input and reference-script info is looked up first in what the builder already holds: selectUtxosFrom UTxOs, inputForEvaluation UTxOs and chainTx outputs. fetchUTxOs is called only for hashes still unknown.
• Concurrent lookups of one hash share one request. A failed or empty response is no longer remembered, so a later complete() retries it instead of failing with "Couldn't find value information".
• fetchCostModels runs alongside the UTxO fetches instead of before them.
• Native (simple) script reference sources are now fetched and completed, including their scriptSize. Before, simpleScriptTxInReference without a hash and size crashed with TypeError: Cannot read properties of undefined (reading 'find'). Native-script inputs given by reference only are now completed too.
• A reference script's hash is derived from scriptRef when the fetcher does not return scriptHash (e.g. ScalusEmulator).
• clone() shares the fetched UTxO cache instead of deep-copying it for every coin-selection callback.

mesh-core-cst (OfflineEvaluatorScalus) and mesh-core-csl (OfflineEvaluator):

• Fetched outputs are reused across evaluateTx calls, since tx outputs never change, and distinct hashes are fetched in parallel. The builder calls the evaluator once per coin-selection round, so each round used to refetch every input, collateral and reference input.
• The cache is capped at 1000 tx hashes per evaluator. Empty or failed responses, and hashes missing a needed output, are dropped so they are refetched.
• OfflineEvaluator no longer appends resolved UTxOs to the caller's additionalUtxos array.

Provider calls, before → after

Measured with ScalusEmulator behind a call-recording proxy (provider-calls.test.ts):

Scenario	Before	After
ADA send, evaluator set	evaluateTx ×1 (×1+k with extra selection rounds)	0
Spend of a chainTx output given as hash/index	fetchUTxOs for an unsubmitted hash, then "Couldn't find value information"	0 fetches, builds and submits
Native script by reference without hash/size	crash	1 fetch per hash, builds and submits
Two script-spend builds with one OfflineEvaluatorScalus	fetchUTxOs ×4	×2
OfflineEvaluator, same tx evaluated twice	fetchUTxOs ×10	×5

No public API changes. The protected getUTxOInfo / queryAllTxInfo keep their signatures.

Validation

npm run build:mesh
cd packages/mesh-transaction && NODE_OPTIONS=--experimental-vm-modules npx jest
cd packages/mesh-core-cst && npm test
cd packages/mesh-scalus-emulator && npm test
cd packages/mesh-core-csl && npx jest test/offline-providers/evaluator.test.ts

• mesh-transaction: 156 tests pass, including the new provider-calls.test.ts. script-tx-fetcher and script-tx-fail fail to load under --experimental-vm-modules on main too, because they import dist through require. Both pass without the flag.
• mesh-core-cst (179 tests) and mesh-scalus-emulator (24 tests) pass.
• In mesh-core-csl/test/offline-providers/evaluator.test.ts the new fetch assertions pass. Four budget assertions in that file already fail on main: expected steps differ after the cost-model update (e.g. 164980381 vs 164973765).