1.4.0

• Breaking: the driver reports the miner fee of the user's own spends as
  fee, read from the transaction on chain: a SwapOut claim's in
  Status::Claiming and Status::Completed, a SwapIn refund's in
  FundState::Refunding and FundState::Refunded, where a batched refund's
  fee is the whole transaction's on every fund it took. Spender carries it
  too. A stored swap still loads, with no fee until a pass reads it again.
• Breaking: Config gains fee_rate_cache_ttl, how long a chain's fee
  estimate is reused, 30 seconds by default, so a sweep reads it once per
  chain rather than once per swap. Zero reads it on every pass, as before.
• Breaking: Config gains server_timeout, 30 seconds by default. A call
  to the server that goes unanswered that long fails as a transport error,
  where the driver used to wait on it without limit.
• Breaking: max_htlc_hold moves from ChainConfig to Config and caps
  a SwapOut hold invoice's min_final_cltv_expiry_delta on its own, 1008
  Lightning blocks by default, where it used to be added to the server's
  dest_csv.
• Breaking: CancelReason gains ClaimWindowClosed { at }. A SwapOut
  with less than claim_margin of the CSV left before its preimage went out
  ends Canceled at once, its payment failing back once the server's CSV
  refund takes the opening from block at. It used to stay Confirming
  until that refund confirmed.
• A SwapOut's Funding.expires_at is no later than the expiry of
  claim_invoice, so it is the time to pay by, and the driver's cancel goes
  out then.
• A SwapOut holds its preimage while the fee estimate is above
  max_fee_rate, so its claim pays the current rate. The swap stays
  Confirming meanwhile.
• cancel answers Requested for a swap already canceled, where it answered
  TooLate. Against a 1.4.0 server, a cancel retried after a lost answer
  gets Requested too.
• run takes a quiet SwapIn up again once the chain's cached fee rate makes
  one of its Dust funds worth refunding, where it used to leave the fund
  Dust until a manual step or refund.
• A SwapOut claim that replaces a pending preimage claim or the server's
  refund goes through the preimage path even once the server's key is known,
  so the replacement pays more in all and is accepted.
• A restored swap's ended_at is the server's transitioned_at only up to
  the time of the restore.
• On wasm32 the driver holds a Web Lock per swap from loading its record to
  saving it, where navigator.locks exists, so tabs and workers over one
  store take turns on a swap. Natively a store still belongs to one process.
• restore_all stops paging at a page with no swap it has not seen whose
  request the wallet signed, where it used to follow the server's has_next
  alone, and reports a swap listed twice once.
• BitcoinEsplora and LiquidEsplora give up on a request after 30 seconds,
  where they used to wait on it without limit; with_timeout sets another
  limit. A ChainClient of your own should bound its requests the same way,
  since the driver waits on a chain call for as long as it takes.
• The driver refunds at most 8 of a SwapIn's funds in one transaction, the
  most valuable first, and the rest once that refund is in a block, so funds
  can stay Refundable while another refund is Refunding. A pass calls
  outspend for as many strays, those in the pending refund first, and for
  none still Locked.
• subscribe yields a swap's handle again when a pass clears its error,
  where a pass that changed nothing else used to stay silent.
• DriverError::retryable is true for an HTTP 429 from the server or
  anything in front of it.