Give linking actions more RAM

[JSGette]

This addresses #628. Current formula sets 50 Mb of RAM for a single linking action which is too low and leads to bazel scheduling too many actions in parallel. This is especially painful in containerized environments (such as kubernetes) where such a behavior can easily cause OOM Killer to kill build pods.

With the new formula, first, default value is slightly higher (256 Mb) and, second, we scale more aggressively with the increasing number of inputs

[meisterT]

This seems to be a drastic increase. Back when we came up with the numbers we did an extensive analysis to come up with the number. I wonder how it changed so signficantly. (cc @wilwell)

[JSGette]

This seems to be a drastic increase. Back when we came up with the numbers we did an extensive analysis to come up with the number. I wonder how it changed so signficantly. (cc @wilwell)

This setting is less relevant for RBE environments where resources are managed at the cluster level (AFAIK, Google mainly relies on RBE so will barely notice this issue if ever).

Our case:
For builds on 6 GB RAM CI workers, it is critical. Bazel immediately claims 1.5 GB for its in-memory cache, but eventually needs 3 GB to handle the full action graph. Since heavy linking actions (C++/Rust) compete for the remaining RAM, the system often hits OOM limits. Using resource_set allows Bazel to throttle concurrent linking actions based on available memory, which is more scalable than manually hard-coding --jobs