Per-unit costs and prices inferred from total should be rounded appropriately

[tesujimath]

With this transaction among others

2025-04-01 * "cost total"
  Assets:Shares:IBM         3 IBM { # 910.00 NZD }
  Assets:NZ:Cash

kiri> limabean --beanfile test-cases/costs-and-prices.beancount

user=> (show (inventory))
Assets:NZ:Cash     -2165.14 NZD
Assets:Shares:IBM      5    IBM  2025-03-01 NZD 300.00
                       3    IBM  2025-04-01 NZD 303.33333333333333333333333333
                       3    IBM  2025-04-05 NZD 302.00
Assets:UK:Cash     -1394.00 GBP

Rounding should be such that tolerance is respected when converting the other way.