Improve performance of dataset_itertator in prepetch mode

[lm-takumi-nakaso]

Improve performance of dataset_itertator in prepetch mode

Change the number of threads from fixed 8 to cpu cores. Share a single process pool with multiple dataset_iterators. Use a single thread per process.

Use loky to achieve this.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[bo-code-review-bot[bot]]

This PR needs Approvals as follows.

• Ownership Approval for / from iizukak, tkng, ruimashita
• Readability Approval for Python from tkng, tsawada, tfujiwar, a-hanamoto

Please choose reviewers and requet reviews!

Click to see how to approve each reviews

You can approve this PR by triggered comments as follows. * Approve all reviews requested to you (readability and ownership) and LGTM review `Approval, LGTM` * Approve all ownership reviews `Ownership Approval` or `OA` * Approve all readability reviews `Readability Approval` or `RA` * Approve specified review targets * Example of Ownership Reviewer of `/`: `Ownership Approval for /` or `OA for /` * Example of Readability Reviewer of `Python`: `Readability Approval for Python` or `RA for Python` * Approve LGTM review `LGTM`

See all trigger comments

Please replace [Target] to review target * Ownership Approval * Ownership Approval for [Target] * OA for [Target] * Ownership Approval * OA * Approval * Readability Approval * Readability Approval for [Target] * RA for [Target] * [Target] Readability Approval * [Target] RA * Readability Approval * RA * Approval * LGTM * LGTM * lgtm

[iizukak]

@lm-takumi-nakaso Thank you for the contribution!! Could you add the speed comparison before and after of this PR?