Fold return collateral dust into the total collateral

[palas]

Context

This is the third of three stacked PRs fixing #1261: the balancing functions
(makeTransactionBodyAutoBalance and estimateBalancedTxBody, in both the traditional and the
experimental API) could build transactions that the ledger rejects at submission time because of
their collateral. Each PR contains a pair of commits — regression tests that fail, then the fix
that makes them pass — and each PR is based on the previous one, so every diff shows only its own
two commits.

🗺️ Map — you are here: 3️⃣

🚪	#1261	the issue: transaction build does not validate min UTxO value of the collateral return output
1️⃣	#1265	remove collateral from transactions without Plutus scripts
2️⃣	#1266	fail on invalid computed collateral instead of building a rejected transaction
3️⃣	(this PR)	fold return collateral dust into the total collateral

This PR: after the previous PR, balancing fails when the ada left over for the return
collateral output is below its minimum UTxO value — even when the collateral is ada-only and a
valid transaction exists. calcReturnAndTotalCollateral (both copies) now folds that ada into
the total collateral and omits the return collateral output instead of failing; the extra ada is
only lost if a Plutus script fails on chain. estimateBalancedTxBody cannot see the content of
the collateral inputs, so it computes collateral under the assumption that they contain no native
tokens — now documented on its totalPotentialCollateral parameters. A transaction built under a
wrong assumption is rejected by the ledger, and tokens are never lost: a valid transaction must
always give the tokens back in the return collateral output. Fixes #1261.

How to trust this PR

• The first commit adds four regression tests (one per balancing function) with an ada-only
  collateral input holding exactly its minimum UTxO value: balancing must succeed, with no return
  collateral output and the whole input as total collateral. On the base branch they fail with
  TxBodyErrorReturnCollateralMinUTxONotMet; the second commit makes them pass.
• The fold only happens when the collateral value is ada-only (isZero on the non-ada part):
  token-carrying collateral keeps failing, and the tests from the previous PR still pass.
• The pre-existing property test prop_calcReturnAndTotalCollateral needs no change: its
  computed-case assertions hold for folded results.

Checklist

• Commit sequence broadly makes sense and commits have useful messages
• New tests are added if needed and existing tests are updated. See Running tests for more details
• Self-reviewed the diff
• Changelog fragment added in .changes/